Search

Image recognition based on deep learning methods has gained remarkable achievements by feeding with abundant training data. Unfortunately, collecting a tremendous amount of annotated images is time-consuming and expensive, especially in grocery product recognition tasks. It is challenging to recognise grocery products accurately when the deep learning model is trained with insufficient data. This paper proposes multi-angle Generative Adversarial Networks (MAGAN), which can generate realistic training images with different angles for data augmentation. Mutual information is employed in the novel GAN to achieve the learning of angles in an unsupervised manner. This paper aims to create training images containing grocery products from different angles, thus improving grocery product recognition accuracy. We first enlarge the fruit dataset by using MAGAN and the state-of-the-art GAN variants. Then, we compare the top-1 accuracy results from CNN classifiers trained with different data augmentation methods. Finally, our experiments demonstrate that the MAGAN exceeds the existing GANs for grocery product recognition tasks, obtaining a significant increase in the accuracy.

Video-based person re-identification plays a critical role in intelligent video surveillance by learning temporal correlations from consecutive video frames. Most existing methods aim to solve the challenging variations of pose, occlusion, backgrounds and so on by using attention mechanism. They almost all draw attention to the occlusion and learn occlusion-invariant video representations by abandoning the occluded area or frames, while the other areas in these frames contain sufficient spatial information and temporal cues. To overcome these drawbacks, this paper proposes a comprehensive attention mechanism covering what, where, and when to pay attention in the discriminative spatial-temporal feature learning, namely What-Where-When Attention Network (W3AN). Concretely, W3AN designs a spatial attention module to focus on pedestrian identity and obvious attributes by the importance estimating layer (What and Where), and a temporal attention module to calculate the frame-level importance (when), which is embedded into a graph attention network to exploit temporal attention features rather than computing weighted average feature for video frames like existing methods. Moreover, the experiments on three widely-recognized datasets demonstrate the effectiveness of our proposed W3AN model and the discussion of major modules elaborates the contributions of this paper.

Recently, intelligent cephalometric landmark detection systems that employ deep convolutional neural networks (DCNNs) have been widely developed. In this paper, we concentrate on two challenging difficulties of DCNNs based cephalometric landmark detection. First, previous DCNN methods usually decreased the resolution of high-resolution cephalometric images. Hence, the trained DCNNs missed the detailed local features of original images. Moreover, the detected landmarks’ locations were recovered using a fixed ratio, so location error was increased after recovery. Second, the performance of cephalometric landmark detection is dependent on DCNNs’ translational invariance because the target landmarks’ locations are influenced by any shift of the cephalometric images. However, modern DCNNs have limited translational invariance, and previous DCNN methods did not take translational invariance into consideration. In this paper, we developed the widely used two-stage framework to resolve the above two challenges. In the first stage, we train a global detection network, which receives the whole images as input, to generate candidate landmarks. In the second stage, we split the input image into patches and train a local refine network, which receives image patches as input, to refine the landmarks’ locations. The proposed local refine network can produce accurate landmarks’ locations because it entirely focuses on the detailed local features of high-resolution cephalometric images. The global detection network and local refine network are structured with translational invariance as their primary consideration. The proposed method achieved robust performance on a private dataset and the public Automatic Cephalometric X-ray Landmark Detection Challenge 2015 dataset.

Machine reading comprehension (MRC) is a challenging task in natural language processing (NLP), which requires machine to determine the corresponding answer to a given passage and question. Whereas, there always exist unanswerable questions in the real world, which poses a new challenge to MRC tasks. Abundant research work has been carried out on the verification mechanism for the answerability of passage-question pairs. However, these researches only focus on its design and implementation, which has limitations in real-world scenarios. Thus, the method proposed in this paper not only verifies the answerability, but also validates and adjusts the predicted answer to obtain an elaborate answer span. Using powerful pre-trained model as encoder block, this paper explores a more comprehensive verification mechanism. Similar to how humans read passages and give answers, we propose a three-stage mechanism called ”Verification for an Elaborate Span” (V4ES): 1) sketchy reading that the model briefly browses the overall information of the passage and question, and then generates an initial answer; 2) intensive reading that it reads the passage and question again, judges the answerability of the question and gives an answer at this stage; 3) verification that it verifies these two answers produced at the previous two stages, and then gives the final prediction. Moreover, the proposed model is evaluated on two MRC challenge datasets: SQuAD2.0 and CMRC2018, and the experiment results show that our model has achieved great improvement compared with the ALBERT and BERT baselines. In conclusion, our proposed verification mechanism has demonstrated its effectiveness through a series of experiments and analysis.

Object detection and semantic segmentation are two fundamental techniques for Intelligent Vehicles (IV) and Advanced Driving Assistance System (ADAS). Motivated by recent studies demonstrating that object detection and semantic segmentation are two highly-correlated tasks, this paper handles the problem of joint object detection and semantic segmentation in traffic scenes. Existing methods perform the joint object detection and semantic segmentation by sharing the same backbone network, but always ignore the interactive connection between the subdividing detection branch and segmentation branch, leading to the insufficient interaction between the two branches. Considering this situation, this paper proposes a joint object detection and semantic segmentation model with the cross-attention and inner-attention mechanisms. The cross-attention mechanism enables to build up the essential interaction between the subdividing detection branch and segmentation branch to fully make use of their correlation. In addition, the inner-attention contributes to strengthening the representations of feature maps in the model. Given an image, an encoder-decoder network is firstly used to extract initial feature maps. Then, the inner-attention mechanism is applied to strengthen the initial feature maps to obtain segmentation feature maps. Subsequently, the cross-attention mechanism utilizes the segmentation feature maps to guide the generation of object detection feature maps. Finally, the semantic segmentation is performed on the segmentation feature maps and object detection is performed on the detection feature maps. In the experiments, two well-known public traffic datasets are used to evaluate our model. Our model achieves the highest performance in comparison with several recently-proposed methods. In addition, some ablation studies are conducted to evaluate the proposed inner-attention and cross-attention mechanisms, and experiment results validate their effectiveness.

This paper focuses on two related sub-tasks of aspect-based sentiment analysis, namely aspect-term extraction and aspect sentiment classification. The former aims to extract aspect-terms from given sentences and the latter aims to identify the sentiment polarity expressed on the extracted terms. Considering the practical application, researchers use more joint methods rather than pipeline methods. However, existing joint methods cannot model the interaction between aspect-terms and the sentence they belong to, or consider the relevance among the sentiments of different aspect-terms. In this paper, a novel end-to-end generative model based on encoder-decoder, namely Joint Term-Sentiment Generator (JTSG), is presented to generate all aspect term-polarity pairs. Specifically, a pre-trained model based encoder is used to encode the sentences, and specially, the decoder generates the start and end position to determine an aspect-term, rather than generate aspect-terms themselves. This new generative method contributes to avoid generating incomplete aspect-terms. Experimental results demonstrate that the proposed approach yields competitive performance on three benchmark datasets.

Brain tumor segmentation using MRI data remains challenging for some reasons. Hence, how to accurately segment the brain tumor is kept as a significant topic in the area of medical image segmentation. To follow the idea of “coarse-to-fine” in practice, this paper proposes an end-to-end two-stage generative adversarial neural network (ToStaGAN) to improve the braintumor is kept as a significant topic in the area of medical image segmentation. To follow the idea of “coarse-to-fine” in practice, this paper proposes an end-to-end two-stage generative adversarial neural network (ToStaGAN) to improve the brain tumor segmentation performance by making full use of semantic information in high-level.In the proposed ToStaGAN, the UNET network is adopted as the “coarse” generation network in the first stage, while a U-Shaped contextual autoencoder (ConEnDer), which consists of the proposed fine-grained extraction module (FEM) and dense skip connection, is proposed as the “fine” generation network in the second stage. To be more specific, the FEM is mainly used to effectively extract “fine-grained” features that plays a crucial role in semantic segmentation with only increasing a few parameters. Furthermore, it has the ability to obtain more diverse and abstract features. And by further utilizing the features extracted from the FEM and the coarse prediction map from the first stage, the ConEnDer holds the great potential to generate the optimized segmentation results. We then evaluate the performance of ToStaGAN on the BARATS 2015. Findings from the evaluations suggest that the proposed two-stage generation network achieves a better performance than the one-stage network; thus, also implying the effectiveness of the proposed ConEnder for “fine” segmentation. A comparative summary also shows that the ToStaGAN achieves a better segmentation performance in comparison to other competing approaches.

Benefiting with the rapid development of communication networks, effective video quality assessment (VQA) models which provide guidance for video transmission and compression technologies are highly demanded. This paper proposes a general-purpose full-reference VQA method combining DenseNet with spatial pyramid pooling and RankNet to not only extract high-level distortion representation and global spatial information of samples but also characterize the temporal correlation among frames. Firstly, the pretrained DenseNet is modified and finetuned to extract high-level features of distorBenefiting with the rapid development of communication networks, effective video quality assessment (VQA) models which provide guidance for video transmission and compression technologies are highly demanded. This paper proposes a general-purpose full-reference VQA method combining DenseNet with spatial pyramid pooling and RankNet to not only extract high-level distortion representation and global spatial information of samples but also characterize the temporal correlation among frames. Firstly, the pretrained DenseNet is modified and finetuned to extract high-level features of distorted videos. Then, spatial pyramid pooling is equipped in the DenseNet module to process flexible inputs with arbitrary spatial resolution. Thus, this kind of input which has the same spatial resolution as the original distorted video is processed by the well-trained DenseNet to generate frame-level quality, which considers the global spatial information of videos directly. Finally, learning to rank is introduced to explore the high-level temporal correlation of distorted videos by taking the RankNet as the temporal pooling function. The experimental results on two public VQA databases show that the proposed algorithm performs consistently with human visual perception.

Person re-identification is to recognize a pedestrian who has been observed at different places in monitoring system, which is done by computer automatically. Important practical significance can be found in this paper on improving the intelligence of monitoring system. There are two challenging problems needed to be solved for person re-identification, which is feature representation and matching. Designing a suitable feature representation method is both challenging and crucial. Ideally, the extracted features should be robust enough to cope with the various situations of the monitoring environment, such as posture difference, illumination changing, and shooting angle difference, etc. There are a lot of challenges for feature matching too, such as small sample size, inter-class confusion, and intra-class variation, etc. In order to meet these challenges, a new person re-identification scheme was put forward in this paper. The following three aspects would be used to describe it. First, an improved BOF feature extraction algorithm based on SURF was put forwarded, SURF algorithm extracts the preliminary feature descriptor and generates visual dictionary, so the influence factors such as illumination and scale invariants can be deal with. And then, covariance descriptor was adopted by the algorithm. It has high robustness, and matching accuracy can be improved too for the case of small sample. Finally, an effective classifier was designed by LIBSVM based on improved BOF algorithm, so the efficiency of the person re-identification algorithm can be improved. The proposed method was compared with the current mainstream algorithm through experiment, and it can be found through the experimental results that it is effective to solve the difficult problems for person re-identification.

Many methods achieve the visual object tracking task with deep learning technologies. As the deep features of different levels contain various semantic information and functions, this paper presents a multi-hierarchy feature aggregation approach to tackle the specific issues in the tracking task, which consists of two aspects. On one hand, this paper integrates the features captured by the offline and online classifiers at the score level, which constructs complementary roles of these classifiers to enhance the stability of classification. Besides, the proposed offline classifier is continuously optimized with different levels of features to reinforce classification constraints. On the other hand, we design a butterfly attention module to promote the capacity of multi-hierarchy feature aggregation in the regression network, which aims to fuse and strengthen the multi-scale features by attending to their spatial information. It can capture more spatial contexts by utilizing the self-attention mechanism during the fusion procedure, and preserve the hierarchy of the features during the strengthening process. Extensive experiments on four public datasets, i.e., VOT2018, OTB100, NFS and LaSOT datasets, demonstrate the effectiveness of the proposed methods.

Context and rationale Intrusion Detection, the ability to detect malware and other attacks, is a crucial aspect to ensure cybersecurity. So is the ability to identify this myriad of attacks. Objective Artificial Neural Networks (as well as other machine learning bio-inspired approaches) are an established and proven method of accurate classification. ANNs are extremely versatile – a wide range of setups can achieve significantly different classification results. The main objective and contribution of this paper is the evaluation of the way the hyperparameters can influence the final classification result. Method and results In this paper, a wide range of ANN setups is put to comparison. We have performed our experiments on two benchmark datasets, namely NSL-KDD and CICIDS2017. Conclusions The most effective arrangement achieves the multi-class classification accuracy of 99.909% on an established benchmark dataset.

Single-object tracking is regarded as a challenging task in computer vision, especially in complex spatio-temporal contexts. The changes in the environment and object deformation make it difficult to track. In the last 10 years, the application of correlation filters and deep learning enhances the performance of trackers to a large extent. This paper summarizes single-object tracking algorithms based on correlation filters and deep learning. Firstly, we explain the definition of single-object tracking and analyze the components of general object tracking algorithms. Secondly, the single-object tracking algorithms proposed in the past decade are summarized according to different categories. Finally, this paper summarizes the achievements and problems of existing algorithms by analyzing experimental results and discusses the development trends.

In the process of Human-robot skill transfer, we require the robot to reproduce the trajectory of teacher and expect that the robot can generalize the learned trajectory. For the trajectory after generalization, we expect that the robot arm can accurately track. However, because the model of the robot can not be accurately obtained, some researchers have proposed using a neural network to approximate the unknown term. The parameters of the traditional RBF neural network are usually selected through the empirical and trial-and-error method, which maybe biased and inefficient. In addition, due to the end-effector of the mechanical arm trajectory will be constantly changing according to the needs of the task, when the neural network of compact set cannot contain the whole input vector, the neural network cannot achieve the ideal approximation effect. In this paper, the broad neural network is used to approximate the unknown terms of the robot. This method can reuse the motion controller that has been learned and complete other motions in the robot operating space without relearning its weight parameters. In this paper, the effectiveness of the proposed method is proved by the ultrasound scanning task.

In this work, we propose a Fractional-order Residual Convolutional Neural Network (FRCNN) for Low-Dose CT (LDCT) denoising. As increasing the dose of radiation is harmful to the patient, how to trade off between reducing the radiation dose and improving the quality of the CT image has become a challenging problem. To this end, this paper proposes a new approach for LDCT image denoising using Convolutional Neural Network (CNN) with Fractional-order Total Variation (FTV) loss, as well as residual learning. Firstly, this paper introduced the FTV loss function for structural details enhancement. Secondly, skip connections were added to optimize the network. Thirdly, extensive experimental analysis was used to evaluate the capacity of this method in suppressing noise and preserving detailed information. The FTV loss can retain essential structural details while suppressing noise, generating high-quality CT images ready for interpretation by radiologists. Compared with state-of-the-art methods, our method obtained better results visually and numerically, especially in structural details preservation. These promising results will significantly improve the usability of LDCT images.

This paper proposes a hybrid architecture based on neural networks, fuzzy systems, and n-uninorms for solving pattern classification problems, termed as ENFS-Uni0 (short for evolving neuro-fuzzy system based on uni-nullneurons). The model can produce knowledge in an on-line (single-pass) and evolving learning context in a particular form of neuro-fuzzy rules representing the dependencies among input features through IF-THEN type relations. The rules antecedents are thereby realized through uni-nullneurons, which are constructed from n-uninorms, leading to the possibility to express both, AND- and OR-connections (and a mixture of these) among the single antecedent parts of a rule (and thus achieving an advanced interpretability aspect of the rules). The neurons’ evolution is done through an extended version of an autonomous data partition method (ADPA). On-line interpretation of the timely evolution of rules is addressed by (i) a concept for tracking the degree of changes of the rules over data stream samples, which may indicate experts/operators how much dynamics is in the process and may be used as a structural active learning component to request operator’s feedback in the case of significant changes and (ii) a concept for updating feature weights incrementally. These weights express the (possibly changing) impact degrees of features on the classification problem: features with low weights can be seen as unimportant and masked out when showing rules to an expert ( → rule length reduction). The rules’ consequents are represented by certainty vectors and are recursively updated by an indicator-based recursive weighted least squares (I-RWLS) approach (one RWLS estimator per class) where the weights are given through the neuron activation levels in order to gain stable local learning. The model proposed in this paper was successfully compared to related hybrid and evolving approaches in the literature for classifying binary and multi-class patterns. The results obtained by the proposed model show an outperformance of the related works in terms of higher accuracy trend lines over time, while offering a high degree of interpretability through coherent neuro-fuzzy rules to solve the classification problems.

Subtle facial expression recognition (SFER) aims to classify facial expressions with very low intensity into corresponding human emotions. Subtle facial expression can be regarded as a special kind of facial expression, whose facial muscle movements are more difficult to capture. In the last decade, various methods have been developed for common facial expression recognition (FER). However, most of them failed to automatically find the most discriminative parts of facial expression and the correlation of muscle movements when human makes facial expression, which makes them unsuitable for SFER. To better solve SFER problem, an attention mechanism based model focusing on salient local regions and their correlations is proposed in this paper. The proposed method: 1) utilizes multiple attention blocks to attend to distinct discriminative regions and extract corresponding local features automatically, 2) a correlation attention module is integrated in the model to extract global correlation feature over the salient regions, and finally 3) fuses the correlation feature and local features in an efficient way for the final facial expression classification. By this way, the useful but subtle local information can be utilized in more detail, and the correlation of different local regions is also extracted. Extensive experiment on the LSEMSW and CK+ datasets shows that the method proposed in this paper achieves superior results, which demonstrates its effectiveness.

The purpose of this paper is to give an overview of various well known medical image registration techniques with a special focus on registration between anatomical and functional medical images. Examples of anatomical medical images (AMI) are Computer Tomography (CT), Magnetic Resonance Images (MRI), X-ray radiographs and ultrasound etc. whereas Positron Emission Tomography (PET), Single Photon Emission Tomography (SPECT) and fMRI are examples of functional medical images (FMI). Irrespective of such types (AMI or FMI), every case can be considered as a medical imaging modality. All these modalities are widely used by clinicians to study the structure/functionality of human body parts for diagnosis and treatment. It is frequently required to combine PET/SPECT with CT/MRI to simultaneous study the metabolic and molecular information received through PET/SPECT with fine anatomical details observed by CT/MRI. This concurrent study will help in the diagnosis and localization of many diseases like cancer, blockage in coronary arteries and brain-related diseases like Parkinson, Alzheimer etc. Further, in many cases, clinicians are required to co-register one or more anatomical images with functional images, for example, ultrasound-guided biopsy fused with PET and MRI. This registration can be done either at the hardware level or at the software level. The introduction of integrated PET-CT machine increases the acceptability of hardware-based registration systems as compared to software-based methods among the medics. One possible reason for this is the lack of validation of results achieved through software-based registration methods. On the other hand, software-based registration methods also have many advantages over the hardware-based registration systems like lesser exposure to radiation and no need for new investment on hardware etc. In this paper, both the methods with their merits and demerits are discussed in detail.

Extreme-learning-machine based auto-encoder (ELM-AE) is regarded as a useful architecture with fast learning speed and general approximation ability, and stacked ELM is used to develop efficient and effective deep learning networks. However, considering features learned from conventional ELM-AEs have issues of weak nonlinear representation ability and random factors in feature projection, this paper proposes an improved ELM-AE architecture which utilize low-rank matrix factorization to learn optimal low-dimensional features. Two superiorities can be obtained compared to conventional ELM-AEs. One is the dimensionality of the hidden layer in ELM-AE could be set arbitrarily, e.g. a higher-dimension hidden layer could lower the random effect in feature learning and enhance features representation ability. The other is enhancing features nonlinear ability, since features are learned directly from the nonlinear outputs of hidden layer. Finally, comparison experiments on numerical and image datasets are implemented in this paper to verify the superior performance of the proposed ELM-AE in this paper.

Deep learning has made substantial progress in crowd density estimation, but in practical application, due to the interference factors such as uneven distribution of crowd and changes in illumination, the existing methods still have large errors in counting. To solve the above problems, a crowd density estimation method based on multichannel dense grouping network is proposed. To solve the problem that it is difficult to extract the multiscale information of the crowd due to the uneven distribution of the crowd, multichannel dense grouping module (McDGM) is designed. In the module, Improved grouping convolution block (IGCB) is dense connected with other layers to obtain different levels of crowd characteristics, so as to avoid the loss of multi-scale information. In addition, the parameters of IGCB are reduced by group convolution. Then, to overcome the change of illumination, a crowd image enhancement algorithm is designed, which makes the image clear by the average pixel value and adjusting the contrast. Finally, to enhance the sensitivity of the network to crowd counting, a new loss function is proposed, which adds counting loss to the previous pixel space loss to improve the accuracy of crowd counting. The algorithm in this paper has been tested on ShanghaiTech and UCF_CC_50 datasets. The test results show that the algorithm in this paper has better statistical accuracy.

In this paper, a neural synchronization of the optimal structure-based group of neural networks is proposed. Asymmetric cryptography is widely used to generate a key amongst two parties and to exchange the key through an insecure channel. However, since the methods that used this strategy, like RSA, have been compromised, new methods for producing a key that can offer security must be discovered. A new group of cryptography known as neural cryptography was created to solve this issue. The primary aim of neural cryptography is to produce a secret key over an unreliable medium. The optimal neural network architecture for creating and defining a secret key between the two authorized individuals is examined in this article. Furthermore, studies into the coordination of a group of neural networks are uncommon. For the design of the public key exchange protocol, synchronization of a cluster of neural networks with Three Layer Tree Parity Machine (TLTPM) is proposed. To calculate the synchronization time, steps taken, and the number of times the attacking neural network could replicate the actions of the two accepted networks, more than 15 million simulations were run. Various parametric experiments have been conducted on the proposed methodology. Simulations of the approach show that it is correct, according to the results of the paper.

With the rapid development of deep learning technology and improvement in computing capability, deep learning has been widely used in the field of hyperspectral image (HSI) classification. In general, deep learning models often contain many trainable parameters and require a massive number of labeled samples to achieve optimal performance. However, in regard to HSI classification, a large number of labeled samples is generally difficult to acquire due to the difficulty and time-consuming nature of manual labeling. Therefore, many research works focus on building a deep learning model for HSI classification with few labeled samples. In this article, we concentrate on this topic and provide a systematic review of the relevant literature. Specifically, the contributions of this paper are twofold. First, the research progress of related methods is categorized according to the learning paradigm, including transfer learning, active learning and few-shot learning. Second, a number of experiments with various state-of-the-art approaches has been carried out, and the results are summarized to reveal the potential research directions. More importantly, it is notable that although there is a vast gap between deep learning models (that usually need sufficient labeled samples) and the HSI scenario with few labeled samples, the issues of small-sample sets can be well characterized by fusion of deep learning methods and related techniques, such as transfer learning and a lightweight model. For reproducibility, the source codes of the methods assessed in the paper can be found at https://github.com/ShuGuoJ/HSI-Classification.git.

Irregular time series data is becoming increasingly prevalent with the growth of multi-sensor systems as well as the continued use of unstructured manual data recording mechanisms. Irregular data and the resulting missing values severely limit the data's ability to be analysed and modelled for classification and forecasting tasks. Often, conventional methods used for handling time series data introduce bias and make strong assumptions on the underlying data generation process, which can lead to poor model predictions. Traditional machine learning and deep learning methods, although at the forefront of data modelling, are at best compromised by irregular time series data sets and fail to model the temporal irregularity of incomplete time series. Gated recurrent neural networks (RNN), such as LSTM and GRU, have had outstanding success in sequential modelling, and have been applied in many application fields, including natural language processing. These models have become an obvious choice for time series modelling and a promising tool for handling irregular time series data. RNNs have a unique ability to be adapted to make effective use of missing value patterns, time intervals and complex temporal dependencies in irregular univariate and multivariate time series data. In this paper, we provide a systematic review of recent studies in which gated recurrent neural networks have been successfully applied to irregular time series data for prediction tasks within several fields, including medical, human activity recognition, traffic monitoring and environmental monitoring. The review highlights the two common approaches for handling irregular time series data: missing value imputation at the data pre-processing stage and modification of algorithms to directly handle missing values in the learning process. Reviewed models are confined to those that can address issues with irregular time series data and does not cover the broader range of models that deal more generally with sequences and regular time series. This paper aims to present the most effective techniques emerging within this branch of research as well as to identify remaining challenges, so that researchers may build upon this platform of work towards further novel techniques for handling irregular time series data.

This paper proposes a solution to translating the visible images into infrared images, which is challenging in computer vision. Our solution belongs to unsupervised learning, which has recently become popular in image-to-image translation. However, existing methods do not produce satisfactory results because (1) most existing methods are mainly used in entertainment scenarios with single scenes and low complexity. The problem solved by this article is more diverse and more complicated. (2) The infrared response of objects depends not only on itself but also on the current environment, and existing methods cannot correlate long-range dependent objects. In this paper, We propose Global Spatial Attention (GSA), which enhances dependence between long-range objects and improves the synthesized image quality. Compared with other methods, GSA can save more space and time. Moreover, we introduce the idea of subspace learning into the neural network to make training more stable. Our method takes unpaired visible images and infrared images for training, which are easy to collect. Experimental results show that our method can generate high-quality infrared images from visible images and outperforms state-of-the-art methods.

Road landmark inventory is becoming an important data product for the maintenance of transport infrastructures. Several commercial sensors are available which include synchronized optical cameras that allowto build 360° panoramic images of the surroundings of the vehicle used for road inspection. This paper is devoted to the analysis of such panorama images,specifically the area that contains themost relevant information. Road lane landmark detection is posed as a two class classification problem that may be solved bymachine learningapproaches, such as Random Forest (RF) and ensembles of Extreme Learning Machines (V-ELM). Besides model parameter selection, a central problem is the construction of a labeled training and validation datasetto cope with the highly uncontrolled conditions of image capture. Besides, human labor cost makes image data labeling a very expensive process. This paper proposes an open ended Active Learning (AL) approach involving a human oraclein the loop who provides the data labeling and can trigger the AL process when detection quality is degraded by the change in imaging conditions. The paper reports encouraging results over a collection of sample images selected from an industrial road landmark inventory operation. As an additional contribution, this paper assesses the ability of AL to overcomesome of the issues raised by highly class imbalanced datasets.

Multi-view clustering has gained soar attention in the field of data mining in recent years. Most multi-view clustering methods are considered to be complete for each view. However, in many applications, some views may contain missing instances, resulting in the incomplete multi-view learning problem. Incomplete views cannot be directly dealt with traditional multi-view clustering methods. In this paper, we propose a One-step multi-view subspace clustering with incomplete views (OMVSC-IV) method based on low-rank matrix factorization to overcome this problem. Specifically, the proposed algorithm uses low-rank matrix factorization to learn a consensus representation matrix, and then combined with the objective function of non-negative embedding and spectral embedding subspace clustering proposed in this paper. The whole process is established into the same objective function for joint optimization and requires no post-processing (e.g., K-means), thus avoiding the defect of being sensitive to initial values. Furthermore, the accuracy, parameter, sensitivity and convergence of the algorithm are studied by experiments. Experimental results on nine data sets show that the proposed algorithm performs superior to some of the latest algorithms.

Depth estimation is a classic task in computer vision, which is of great significance for many applications such as augmented reality, target tracking and autonomous driving. Traditional monocular depth estimation methods are based on depth cues for depth prediction with strict requirements, e.g. shape-from-focus/ defocus methods require low depth of field on the scenes and images. Recently, a large body of deep learning methods have been proposed and has shown great promise in handling the traditional ill-posed problem. This paper aims to review the state-of-the-art development in deep learning-based monocular depth estimation. We give an overview of published papers between 2014 and 2020 in terms of training manners and task types. We firstly summarize the deep learning models for monocular depth estimation. Secondly, we categorize various deep learning-based methods in monocular depth estimation. Thirdly, we introduce the publicly available dataset and the evaluation metrics. And we also analysis the properties of these methods and compare their performance. Finally, we highlight the challenges in order to inform the future research directions.

With the advent of smart phones, the popularity of free Android applications has risen rapidly. This has led to malicious Android apps being involuntarily installed, which violate the user privacy or conduct attack. Malware detection on Android platforms therefore is a growing concern because of the undesirable similarity between malicious behavior and benign behavior, which can lead to slow detection, and allow compromises to persist for comparatively long periods of time in infected phones. The contributions of this paper are first a multiple dimensional, kernel feature-based framework and feature weight-based detection (WBD) designed to categorize and comprehend the characteristics of Android malware and benign apps. Furthermore, our software agent is orchestrated and implemented for the data collection and storage to scan thousands of benign and malicious apps automatically. We examine 112 kernel attributes of executing the task data structure in the Android system and evaluate the detection accuracy with a number of datasets of various dimensions. We find that memory- and signal-related features contribute to more precise classification than schedule-related and other descriptors of task states listed in our paper. Particularly, memory-related features provide fine-grain classification policies for preserving higher classification precision than the signal-related and others. Furthermore, we study and evaluate 80 newly infected attributes of the Android kernel task structure, prioritizing the 70 features of most significance based on dimensional reduction to optimize the efficiency of high-dimensional classification. Our second contribution is that our experiments demonstrate that, as compared to existing techniques with a short list of task structure features (16 or 32 features), our method can achieve 94%-98% accuracy and 2%–7% false positive rate, while detecting malware apps with reduced-dimensional features that adequately abbreviate online malware detections and advance offline malware inspections.

Recently, due to its widespread applications in public safety, anomaly detection in crowd scenes has become a hot topic. Some deep-learning-based methods attain significant achievements in this field. Nevertheless, most of them suffer from over-fitting to some extent because of scarce data, which are usually abrupt and low-frequency in the real world. To remedy the above problem, this paper firstly develops a synthetic anomaly event generating system, which could simulate typical specific abnormal events. By utilizing this system, a large synthetic, diverse anomaly event dataset is built, which contains 2,149 video sequences. After getting the dataset, a 3D CNN is designed to detect the abnormal types at the video level. However, we find that there are obvious domain differences (also named as “domain gap/shifts”) between synthetic videos and real-world data, which results in performance degradation when applying the model to the real world. Thus, this paper further proposes a cyclic 3D GAN for domain adaption to reduce the domain gap, which translates the synthetic data to the photorealistic video sequences. Then the detection model is trained on the translated data and it can perform well in the real data. Experimental results illustrate that the proposed method outperforms these baselines for the domain adaptation anomaly detection.

Clustering is used in many applicative fields to summarize information into a small number of groups. Motivated by behavioral extraction issues from urban data, the interest of this paper is to propose a classification method that allows modeling the evolution of cluster profiles over time while considering common regressive effects. The parameters of the proposed model are estimated using variational approximation because maximum likelihood estimation is not suitable in this case. The ability of the model to estimate parameters is evaluated using various simulated data and compared with two other models.

The accurate segmentation of cerebral vessels from time-of-flight magnetic resonance angiography (TOF-MRA) data is crucial for the diagnosis and treatment of cerebrovascular diseases. However, cerebrovascular segmentation remains challenging due to the tiny and complex structure of the vessels. This paper proposed a GAN-based deep-based method that maps the TOF-MRA data to 3D cerebral vessels. To improve the capability of feature expression, the proposed network integrates squeeze-and-excite blocks into a Vnet structure to extract and accumulate features at different scales. In addition, we introduce adversarial loss for network training, which can help the model to learn the distribution of the cerebral vessels. Besides, we introduce island number and minimum diameter to evaluate the spatial consistency and recognized minimal vessel of the cerebrovascular segmentation results. The proposed method is validated on 102 TOF-MRA cases with manually annotated labels. The experimental results indicate that our method can identify more tiny cerebral vessels meanwhile maintaining the spatial consistency of the cerebral vessel structure. Besides, our GAN-based cerebrovascular segmentation method reaches an average dice similar coefficient of 89.89 % , outperforming other state-of-the-art segmentation methods.

Online transfer learning (OTL) has attracted much attention in recent years. It is designed to handle the transfer learning tasks, where the data of the target domain isn’t available in advance but may arrive in an online manner, which may be a more realistic scenario in practice. However, there typically are two limitations of existing OTL algorithms. 1) Existing OTL algorithms are based on shallow online learning models (SOLMs), e.g., linear or kernel models. Due to this limitation of SOLMs they cannot effectively learn complex nonlinear functions in complicated application and the OTL algorithms based on SOLMs cannot either. 2) Existing algorithms only utilize the latest arrived instance to adjust the model. In this way, the previously arrived instances are not utilized. It may be better to utilize the previously arrived instances as well. In this paper, to overcome the abovementioned two limitations, a buffered online transfer learning (BOTL) algorithm is proposed. In the proposed BOTL algorithm, the learner is designed as a deep learning model, referred to as Online Hedge Neural Network (OHNN). In order to enable the OHNN to be effectively learned in an online manner, we propose a buffered online learning framework that utilizes several previously arrived instances to assist learning. Further, to enhance the performance of the OHNN, a model learned in the source domain is transferred to the target domain. The regret bound of the proposed BOTL algorithm is analyzed theoretically. Experimental results on realistic datasets illustrate that the proposed BOTL algorithm can achieve lower mistake rate than the algorithms compared.

With the rapid development of Internet technology and the explosive growth of digital text, opinion mining has become one of the important research hotspots in the field of natural language processing (NLP). In recent years, neural network based deep learning algorithms have been applied in the field of opinion mining. Considering the relation between temporal and spatial dimensions of text data and the characteristics of natural language itself, traditional deep learning algorithms cannot be comprehensive in the processing of fully feature extraction. In this paper, we propose a new deep learning framework for opinion mining, which includes a temporal feature extraction layer that consists of two layers of bidirectional simple recurrent unit (Bi-SRU) networks extracting features at the word and grammar levels; a semantic feature extraction layer that mainly contains a multi-head attention module; a spatial feature extraction layer with dilated convolution that is used to extract opinion preference features. The Internet movie database (IMDb) is used to verify the performance of the proposed framework. The experiment results show that the proposed framework can effectively improve the classification accuracy, whose performance is better than that of the compared algorithms.

Video temporal segmentation plays a vital role in video analysis since many higher-level computer vision tasks rely on it. Some recent efforts have been dedicated to generating temporal action proposals for long and untrimmed videos, which requires methods to generate accurate boundaries for video semantics. In this paper, we propose a novel and efficient Temporal Distribution Network (TDN), to model the long-term distribution of video semantic units (video dictionary). Firstly, we encode the semantics and context relations of video segments with a boundary-specified video embedding method. Then based on temporal convolutional layers, we design a Temporal Distribution Network (TDN) enumerating all the possible temporal locations in one pass and generating proposals that have high action confidence scores by capturing the long-term distributions of video semantics. We validate our method on temporal action proposal generation tasks and action detection tasks. Experimental results on two benchmark datasets, THUMOS14 and ActivityNet-1.3, show that the proposed method can significantly outperform the state-of-the-art approaches. Our model could obtain high-quality action proposals with a much faster speed.

This paper shows a computational model that classifies Cybergrooming attacks in the context of COP (child online protection) using Natural Language Processing (NLP) and Convolutional Neural Networks (CNN). The model predicts a high number of false positives, therefore low precision and F-score, but a high accuracy. In this issue, where the number of messages in the context of grooming are so low compared to the number of conversations and messages from other contexts, it can be concluded that is a very consistent and useful result as it captures a high number of true positives, considering that the classifier works for messages. Performing the training of machine learning algorithms with neural networks, semantic analysis and NLP, allows approximate representation of knowledge contributing to discovery of pseudo-intelligent information in these environments and reducing human intervention for characterization of underlying abnormal behavior and detecting messages that potentially represent these attacks.

Human has excellent motor capability and performance in completing various manipulation tasks. During some tasks such as tightening/loosening a screw with a screwdriver, the motion is accompanied by force exertion to the environment (that is, constrained motion). To obtain natural human-robot interaction (HRI) as human interacts/collaborates with the environment, interpreting the human’s intention in a way of motion and interaction force is meaningful for carrying out constrained tasks. This paper proposes a long-short term memory (LSTM) network-based decoding method for the simultaneous estimation of human motion and interactive force from surface electromyography (sEMG) signals. The surface EMG recorded from the muscles of forearm is used to decode human’s motion intention. In order to extract smooth features from non-stationary sEMG signals, Bayesian filter is applied instead of traditional time-domain feature extraction method. From the real-time experiments on eight subjects, the LSTM-based decoding method represents high accuracy of motion estimates (91.7%) and force estimates (96.1%) despite of the existence of muscle coupling and non-stationary mapping during such constrained tasks. It indicates that the estimated motion and interactive force can be further applied for HRI in accomplishing such constrained tasks.

Convolutional Neural Networks (CNNs) have won tremendous success in various applications, such as image recognition and natural language processing. Due to the computing-intensive and memory-intensive features of CNN models, it is challenging to deploy them on devices with limited resources and tight power budgets. To address this limitation, pruning, quantization, weights sharing, and other methods have been proposed as efficient solutions to compress CNN models. However, these works usually do not optimize the hardware design when compressing the CNNs. In this paper, we introduce Conv-Inheritance, a hardware-efficient compression method for CNNs by reducing the number of convolution operations to simultaneously reduce the inference time and hardware computing resources on chip of CNN edge devices. We also develop an extended version, called adaptive Conv-Inheritance, to further improve the inference performance by considering the similarity between pixels. We applied our Conv-Inheritance method to compressing various CNN models and performed extensive experiments on different image datasets. Experimental results demonstrate that Conv-Inheritance is both hardware-efficient and time-efficient on edge devices. Besides, Conv-Inheritance is a plug-and-play compression method for CNNs, and we can run Conv-Inheritance on top of other compression methods without any conflicts.

Machine Learning has emerged in the last years as the main solution to many of nowadays’ data-based decision problems. However, while new and more powerful algorithms and the increasing availability of computational resources contributed to a widespread use of Machine Learning, significant challenges still remain. Two of the most significant nowadays are the need to explain a model’s predictions, and the significant costs of training and re-training models, especially with large datasets or in streaming scenarios. In this paper we address both issues by proposing an approach we deem predictive and user-centric. It is predictive in the sense that it estimates the benefit of re-training a model with new data, and it is user-centric in the sense that it implements an explainable interface that produces interpretable explanations that accompany predictions. The former allows to reduce necessary resources (e.g. time, costs) spent on retraining models when no improvements are expected, while the latter allows for human users to have additional information to support decision-making. We validate the proposed approach with a group of public datasets and present a real application scenario.

Despite many decades of research on high-performance Deep Neural Network (DNN) accelerators, their massive computational demand still requires resource-efficient, optimized and parallel architecture for computational acceleration. Contemporary hardware implementations of DNNs face the burden of excess area requirement due to resource-intensive elements such as multipliers and non-linear Activation Functions (AFs). This paper proposes DNN with reused hardware-costly AF by multiplexing data using shift-register. The on-chip quantized log 2 based memory addressing with an optimized technique is used to access input features, weights, and biases. This way the external memory bandwidth requirement is reduced and dynamically adjusted for DNNs. Further, high-throughput and resource-efficient memory elements for sigmoid activation function are extracted using the Taylor series and its order expansion have been tuned for better test accuracy. The performance is validated and compared with previous work for the MNIST dataset. Besides, the digital design of AF is synthesized at 45 nm technology node and physical parameters are compared with previous work. The proposed hardware reused architecture is verified for neural network 16:16:10:4 using 8-bit dynamic fixed-point arithmetic and implemented on Xilinx Zynq xc7z010clg400 SoC using 100 MHz clock. The implemented architecture uses 25% less hardware resources and consumes 12% less power without performance loss, compared to other state-of-the-art implementations, as lower hardware resources and power consumption are especially important for increasingly important edge computing solutions.

With the rapid development of cloud computing, Internet of things and artificial intelligence, human–computer interaction (HCI) is playing an increasingly important role in the daily life. As an important component of HCI, hand gesture recognition (HGR) system is usually combined with edge computing server, utilizing machine learning, including neural network, decision tree, integrated learning, to achieve low latency and high reliability service. High precision HGR with low computational complexity is prerequisite for the commercialization of gesture recognition. Therefore, this paper proposed a high-precision parameter correction algorithm based on the established scattered-point model and the outlier detection scheme, and a recognition algorithm with multivariable decision tree is then presented for the dynamic hand gestures. The experimental results show that the proposed algorithms can improve the recognition accuracy and effectively reduce the running time, which is conducive to algorithm transplantation and model deployment in edge servers.

Person re-identification (re-ID) aims at matching a person-of-interest across various non-overlap cameras with distinguished visual appearance variances. Pre-existing research methods mainly employ deep neural models to train large-scale person re-ID datasets, achieving good performance. However, these methods are primarily deployed only on visual data, which can be easily influenced by the environment variances (e.g., viewpoints, poses, and illuminations). In this paper, we propose an adaptive multi-task learning (MTL) scheme for cross domain and modal person re-ID. It can effectively utilize the visual and language information from multiple datasets for improving learning performance. Comprehensive experiments are also conducted on the widely-used person re-ID datasets, i.e., Market-1501 and DukeMTMC-reID, validating the effectiveness of the proposed method. It can model the domain difference and the relationship between the vision and language modalities and achieve state-of-the-art performance. The source code of our proposed method will be available at (https://github.com/emdata-ailab/Multitask_Learning_ReID).

In this paper, we propose the co-attention dictionary network (CODNet) for weakly-supervised semantic segmentation using only image-level class labels. The CODNet model exploits extra semantic information by jointly leveraging a pair of samples with common semantics through co-attention rather than processing them independently. The inter-sample similarities of spatially distributed deep features are computed to merge reference features through non-local connections. To discover similar patterns regardless of appearance variations, we propose to extract image representations by equipping the neural networks with dictionary learning which provides the universal basis elements for different images. Based on the CODNet model, we propose a multi-reference class activation map (MR-CAM) algorithm which generates semantic segmentation masks for a target image by jointly merging semantic cues from multiple reference images. Experimental results on the PASCAL VOC 2012 and MSCOCO benchmark datasets for weakly-supervised semantic segmentation show that the proposed algorithm performs favorably against the state-of-the-art methods.

As one of the most complex social problems around the world, crime may bring the risk of dying or losing property to the public if not handled properly. Crime prediction which aims at predicting crime incidents before they happen is of great importance to fight against crime. Previous studies are concerned primarily with day-level crime prediction and have certain limitations on modeling complex spatial–temporal-categorical dependency contained in the criminal activities as well as utilizing external factors to facilitate the forecast. In this paper, we develop a novel Neural Attentive framework for Hour-level Crime prediction (NAHC) to cope with these challenges. Specifically, we first adopt the priori knowledge-based data enhancement strategy to alleviate the zero-inflated issue raised in hour-level settings. Then, multi-graph convolutional networks are applied to capture spatial dependency from different aspects. After that, we integrate gated recurrent units with a temporal attention mechanism to jointly address temporal dependency and capture time-sensitive external factors. A categorical attention mechanism is proposed for dealing with categorical dependency and finally a fully connected network is utilized to generate the final prediction results. Extensive experiments on two real-world crime datasets demonstrate the effectiveness of our framework over the state-of-the-art comparing methods.

Occluded person re-identification is a challenging task since the available data often suffers from information incompleteness and spatial misalignment. Most state-of-the-art occluded models rely on the external model to provide additional semantic information. However, for the time being, external models, such as the human parsing model and the pose estimation model cannot provide accurate semantic information under a complex occlusion environment and may introduce errors to the Re-ID model instead. In this paper, we propose an occluded person Re-ID model that mines the latent recognizable information of the person image itself, without the help of external models. Feature/Data uncertainty can reduce the influence of noisy samples in datasets and has been discussed in person Re-ID and face recognition, we extend the uncertainty to the micro feature level, and propose the spatial-wise and channel-wise feature uncertainty to constantly refine the features in the spatial domain and the channel domain respectively during feature construction by weakening the influence of noise features. Extensive experiments on the occluded datasets and holistic datasets have proved the effectiveness of our proposed methods.

Neural architecture search (NAS) is to automatically design task-specific neural architectures, whose performance has already surpassed those of many manually designed neural networks. Existing NAS techniques focus on searching for the neural architecture and training the optimal network weights from the scratch. Nevertheless, it could be essential to study how to tune a given neural architecture instead of producing a completely new neural architecture in some scenarios, which may lead to a more optimal solution by combining human experience and the advantages of the machine’s automatic searching. This paper proposes to learn to tune the architectures at hand to achieve better performance. The proposed Neural Architecture Tuning (NAT) algorithm trains a deep Q-network to tune neural architectures given a random architecture so that we can achieve better performance on a reduced space. We then apply adversarial autoencoder to make the learned policy be generalized to a different searching space in real-world applications. The proposed algorithm is evaluated on the NAS-Bench-101 dataset. The results indicate that our NAT framework can achieve state-of-the-art performance on the NAS-Bench-101 benchmark, and the learned policy can be adapted to a different search space while maintaining the performance.

Edge computing and artificial intelligence (AI), especially deep learning for nowadays, are gradually intersecting to build a novel system, called edge intelligence. However, the development of edge intelligence systems encounters some challenges, and one of these challenges is the \textit{computational gap} between computation-intensive deep learning algorithms and less-capable edge systems. Due to the computational gap, many edge intelligence systems cannot meet the expected performance requirements. To bridge the gap, a plethora of deep learning techniques and optimization methods are proposed in the past years: light-weight deep learning models, network compression, and efficient neural architecture search. Although some reviews or surveys have partially covered this large body of literature, we lack a systematic and comprehensive review to discuss all aspects of these deep learning techniques which are critical for edge intelligence implementation. As various and diverse methods which are applicable to edge systems are proposed intensively, a holistic review would enable edge computing engineers and community to know the state-of-the-art deep learning techniques which are instrumental for edge intelligence and to facilitate the development of edge intelligence systems. This paper surveys the representative and latest deep learning techniques that are useful for edge intelligence systems, including hand-crafted models, model compression, hardware-aware neural architecture search and adaptive deep learning models. Finally, based on observations and simple experiments we conducted, we discuss some future directions. Submitted/Accepted version

In autonomous vehicle technology, environmental perception is a very important part of the whole system. The robustness and accuracy of the perception determine the performance of the whole system. Compared with the camera, lidar can generate point cloud data which can provide accurate 3D information better. At present, more researches focus on environmental perception through point cloud. However, the point cloud has the characteristics of sparsity, unordered and non-uniform distribution, its processing method is also different from the general image-based method. In this paper, we proposed a method of semantic segmentation of point cloud which is realized by convolutional neural network. Through spherical mapping, the scattered point cloud is transformed into 2D spherical map with dense and uniform distribution, which is convenient for the processing of the network. To reduce the loss of point information, the spatial position, reflection intensity information and angle value of point are preserved during spherical mapping. Our work includes a lightweight convolutional neural network that fuses point cloud and image information. In addition, a spatial module is added to supplement the loss of spatial information and improve the performance of segmentation. Finally, a series of comparative experiments and performance evaluation are carried out, which show that the proposed method is not only much better than the original structure, but also better than another image and lidar fusion network in the segmentation of each category.

In this paper, a double sparse low-rank decomposition method is proposed to defect detection for complex irregular printed fabrics. Firstly, a low rank decomposition model with double sparsity is established by taking the sparse components as the printing template prior and the difference between the defective printed fabric graph and the template fabric graph as the defect prior. Secondly, the decomposition is guided by printing prior and defect for the double sparse low rank decomposition model to obtain the saliency map of defects. Furthermore, the defect map is obtained by binarising the defect’s saliency map using the optimal threshold segmentation. Finally, the simulation results are compared with the four existing methods. The results show that the proposed algorithm can effectively detect defects in three types of irregular printed fabrics, such as small-size print, medium-size print, and complex-distribution print. The valid positive rate is 89.29%, the false positive rate is 0.85%, and the positive predictive value is 86.21%. Comparison results show that the proposed algorithm retains the shape details of the defect better than the other four algorithms, and the detection time is 12.49% less than the current optimal PN-RPCA algorithm.

With the development of big data and artificial intelligence, distributed optimization has emerged as an indispensable tool for solving large-scale problems. In particular, the multi-agent system based on distributed information processing can be elaborately designed for distributed optimization, in which the agents collaboratively minimize a global objective function made up of a sum of local objective cost functions subject to some local and/or global constraints. Inspired by the applications involving resource allocation, machine learning, power systems, sensor networks and cloud computing, a variety of distributed optimization models and algorithms have been investigated and developed. The optimization models include unconstrained and constrained problems in continuous and discontinuous systems with undirected and directed communication topology graphs. The constraints include bounded constraint, separable and inseparable equality and inequality constraints. Meanwhile, in distributed algorithms, every agent executes its local computation and updating on basis of its own data information and that exchanging with its neighboring agents by means of the underlying communication networks, in order to deal with the optimization problems in a distributed way. This paper is designed to provide a comprehensive overview of extant distributed models and algorithms for distributed optimization.

Recently, face hallucination, also termed face super-resolution (SR), has been widely studied and achieved significant progress. The algorithm based on manifold learning is one of the primary methods for SR. However, when recovering the high-resolution (HR) counterparts from extremely low-resolution (LR) images, the basic assumption of manifold consistence between LR/HR spaces is doubtful and vulnerable. To address this issue, some algorithms usually employ the cascaded models in line with the inherent magnification factors, such as 2, 4 and 8, to maintain the manifold consistence. As a simple cascade mechanism, the inherent factors cannot ensure the optimal performance without the relationship between the manifold relevance and the down-sampling scale. In this paper, we explore the relevance with the groups of gradually down-sampled training sets and divide the scales into different classes for robust manifold consistence. And then, to map the optimal coefficients from LR spaces to HR target ones better, we introduce the weight-mapping neighbour embedding model. Qualitative and quantitative evaluations demonstrate that the enhancing of the manifold relevance can promote effective face hallucination. Based on the weight-mapping and scale clustering, our algorithm achieves better results compared with the state-of-the-art methods.

In this paper, we proposed a more robust supervised hashing framework based on the Cauchy loss function and Supervised Discrete Hashing (SDH) called Robust Supervised Discrete Hashing (RSDH), which can learn a robust subspace consisted of binary codes. The Cauchy loss is used to measure the error between the label matrix and the product of the decomposed matrices. RSDH can not only reduce the outliers and noise of the hashing codes, but also achieve the more satisfactory retrieval effect. Image retrieval experiments demonstrate that RSDH performs better than the other hashing methods.