Search

Word spotting in handwritten documents is challenging due to the high intra-class and inter-class variability of handwritten forms. This paper addresses the word spotting problem in the segmentation and the training scenarios. Overall, this paper makes the following three contributions: (1) a new word text representation, called the Pyramid of Bidirectional Character Sequences (PBCS), which can solve both the word spotting problem and the word recognition problem. The use of the PBCS representation allows trained models to identify the character subsequences shared by words. Thus, words that are not seen during training can be represented and spotted. In addition, the PBCS representation encodes word texts redundantly, allowing for word discrimination. (2) A binary classification modeling of the word spotting problem in the difference space between representations, where spotting of non-vocabulary words is more efficient. Finally, (3) a new deep neural network architecture that combines the strengths of convolutional layers and transformers. We evaluated our solution on IAM and RIEMS datasets and showed that it outperforms recent state-of-the-art methods in the query-by-example scenario. • The Pyramid of Bidirectional Character Sequences (PBCS) word text representation. • Word spotting task as a binary classification problem. • Combing the strengths of convolutional layers and transformers. [ABSTRACT FROM AUTHOR]

• Deep and handcrafted features are extracted from 2D depth images and 3D scans respectively. • The extracted features are embedded into a covariance pooling for the dimensionality reduction. • Deep covariance learning is carried out using two additional layers to enhance their discrimination power. • The Gaussian kernel is used to generalize the radial basis function to the manifold space for SVM classification. • High classification performance is achieved on two challenging datasets. Facial emotion recognition (FER) methods have been proposed mainly using 2D images. These methods suffer from many problems caused by the difficult conditions of unconstrained environments such as light conditions and view variations. In this paper, we aim to recognize the emotional facial expressions independently of their identity using the 3D data and 2D depth images. Since the 3D FER is a very fine-grained recognition task, mapping the 3D images into 2D depth images may lack some geometric characteristics of the expressive face and decay the FER performance. Convolutional Neural Networks (CNN), however, have been successfully applied to the 2D depth images and improved handcrafted-based methods in computer vision and pattern recognition applications. For this reason, we combine in this paper two types of features; handcrafted and deep learning features and prove their complementarity for 3D FER. Favorably, covariance descriptors have proven a very good ability to combine features from different types into a compact representation. Therefore, we propose to use the covariance matrices of features (handcrafted and deep ones), instead of the features independently. Since covariance matrices belong to one of the manifold space types, formed by SPD (Symmetric Positive Definite) matrices, we mainly focus on the generalization of the RBF kernel to the manifold space for 3D FER using a supervised SVM classification. The achieved performance of the proposed method on the Bosphorus and BU-3DFE datasets outperforms similar state-of-the-arts. [ABSTRACT FROM AUTHOR]

• An automatic energy leakage detection for district heating systems based on UAV Infrared (IR) images is presented. • A deep learning-based model and 8 different ML classifiers have been developed and evaluated for detection of leakages. • A region extraction algorithm has been incorporated to extract potential leakages enabling the ML analysis. • The proposed ML methods have been applied to ca. 13.4 million image patches and achieved an actual detection rate of 98.6%. In this paper, we propose automatic energy leakage detection in underground pipes of district heating systems based on Infrared (IR) images, captured by an Unmanned Aerial Vehicle (UAV). Hot water or steam is distributed to homes and industries through underground pipes from a central power plan. Leakages in underground pipes pose a very common problem, which can occur for many reasons, e.g. unprofessional installation and end of service life. Potentially, a leakage remains undiscovered for a very long period of time. Therefore, it is of great interest for power supply companies to monitor district heating networks to identify leakages. In this paper, the original IR images are captured in a 16 bit format by a UAV. On ground, potential leakages are extracted using a region extraction algorithm. Thereafter a Convolutional Neural Network (CNN) as well as eight conventional Machine Learning (ML) classifiers are applied on these regions to classify whether or not it is a leakage. In total, twelve UAV sequences are captured at different cities in Denmark. Based on these, around 13.4 million samples of image patches of district heating systems are extracted. Eleven sequences are used for training and the remaining one for testing. This was performed on all splits in the leave-one-out testing. The deep learning CNN achieved an average weighted accuracy of 0.872 with a false positive and negative rate of 12.7 % and 10.4 %, respectively. This CNN model detected around 98.6 % of the true leakages. In comparison, conventional ML classifiers, i.e. Adaboost (AB), Random Forest (RF), etc. provide lower average weighted accuracy, but on the other hand they require less computational resources. We have compared our method with a state-of-art method and the result shows that the proposed method is very competitive. [ABSTRACT FROM AUTHOR]

• A new approach for PD detection based on CNN deep features and i-vector features. • The impact of effective use of dynamic features for PD classification. • The significance of CNN deep features and i-vector features in terms of PD detection. Parkinson's disease (PD) corresponds to one of the most common neurological diseases in the world, which is mainly manifested by motor, cognitive, and language disorders. The change in the patient's voice is one of the most striking clinical signs and proves to be an element of interest to support the diagnosis and the assessment of PD. In this research paper, a new approach based on speech signal analysis is set forward to automatically detect Parkinson's disease. The approach evaluates two learning techniques, namely Support Vector Machines (SVM) and Convolutional Neural Networks (CNN), to classify data obtained from speech tasks. Two input data,i.e., the raw speech signal's values and the i-vector features of dimensions 100, 200 and 300 are extracted in this study. Eventually, an evaluation step is undertaken through the use of five evaluation metrics which are accuracy, precision, recall/sensitivity, specificity and f -score. The most pertinent obtained results for a test dataset composed of 28 participants are recorded as follows: an accuracy of 100%, precision of 0.99, recall/sensitivity of 0.98, specificity of 0.96 and f -score of 0.98. These findings corroborate that our approach can help in terms of PD detection. [ABSTRACT FROM AUTHOR]

• We propose a rotation invariant Gabor convolutional neural network (RIGCN). • We learn Gabor-guided convolutional features in the Siamese network architecture. • We explore a convolutional fusion operator to obtain rotation invariant features. • We show the effectiveness of RIGCN for rotation invariant image classification. Gabor filters have been recently integrated with deep convolutional neural networks to learn better features with fewer model parameters. However, during feature learning the rotation invariance is not well addressed. In this paper, we propose a rotation invariant Gabor convolutional neural network (RIGCN) for image classification. First, we transform each input image to generate multiple rotated image instances and feed them into a weight-sharing Siamese network architecture to learn Gabor-guided deep convolutional features. Then, we compute the maximum and average feature responses from all the rotated instances of the same input image and send them into a convolutional fusion module to obtain a rotation invariant image representation. Finally, we use the cross-entropy loss for image classification. In our method, the use of Siamese architecture enables us to obtain rotation invariant features from rotated image instances. The use of convolutional fusion operator enables us to obtain richer statistical features with high efficiency. Experimental results on several benchmark datasets demonstrate the effectiveness of RIGCN for rotation invariant image classification. [ABSTRACT FROM AUTHOR]

• We first exploit a deep framework with pattern theory to learn continuous temporal embedding. • We design a configuration proposal module to remove background information in video. • We explicitly formulate an incremental Viterbi algorithm for pattern theory. • The experimental results show that our CTE-PT overforms state-of-the-art methods. Visual Question Answering (VQA) is a challenging task in artificial intelligence and has received increasing attention from both the computer vision and the natural language processing communities. Joint embedding learning for it suffers from the background noise of images and text. Learning continuous temporal embedding is a potential solution to extract both visual elements and textual elements with a proper temporal structure. In this paper, we address a continuous temporal embedding model based on pattern theory (CTE-PT) and fully express the atomism and combinatory based on Grenander's pattern theory. First, we generate atomic actions from videos and note them as generators, which reflects the atomism in pattern theory. Second, we design a CTE-PT model to discover the discriminative configuration of videos, which reflects the combinatory of atomic actions in pattern theory. In the CTE-PT model, we design the configuration proposal module to remove some background information initially, and the configuration interpretation module to minimize the interpretive energy of continuous temporal embedding. We estimate the energy by each pair in this embedding sequence and optimize it based on the Viterbi algorithm. The experimental results show that our CTE-PT model outperforms the baseline C3D+LSTM model on Olympic Sports, UCF 101, and HMDB51 datasets, which proves the effectiveness of mining the common continuous temporal embedding as the class-specific configuration for activity discriminator. [Display omitted] [ABSTRACT FROM AUTHOR]

• A spatio-temporal Siamese neural network is proposed for 3D signature recognition. • The system uses one shot learning to learn from small amounts of training data. • The proposed architecture outperforms conventional Siamese networks. • The ST-SNN outperforms current recognition techniques on raw 3D signature data. Signature forgery is at the centre of several fraudulent activities and legal battles. The introduction of 3D signatures, the virtual signing of ones name in the air, has the potential to restrict forgers due to the absence of visual cues that can be easily copied. Existing 3D signature recognition approaches, however, have not leveraged the inherent spatial and temporal information, making it difficult to handle the diminished separability and reproducibility of these signatures. In this paper, we propose a novel spatio-temporal adaptation of the Siamese Neural Network, wherein one branch extracts spatial features using a 1D Convolutional Neural Network (CNN) while the other processes the input in the temporal domain using Long Short-Term Memory networks (LSTMs). Unlike conventional deep learning networks, Siamese networks are an application of One-Shot Learning so as to learn from a small amount of data as is often the case in real life problems. They employ a distance metric that is forced to be small for like samples (signatures from the same person), and large for different samples (from different persons). The proposed approach, termed ST-SNN, is compared to other baseline classification architectures, and demonstrated using a publicly available biometric 3D signature benchmark dataset, yielding True Positive Rate (TPR) of 94.63% with 4.1% False Acceptance Rate (FAR). [ABSTRACT FROM AUTHOR]

• An end to end framework for clothes image captioning is developed based on attribute detection and visual attention. • The attribute detection module is developed to provide the encoded feature of clothes with more attribute information. • An attention gate is adopted to better characterize the similarity between image portion and captioning description. • Extensive simulations based on real-world data are conducted to verify the effectiveness of the proposed method. Fashion is a multi-billion-dollar industry, which is directly related to social, cultural, and economic implications in the real world. While computer vision has demonstrated remarkable success in the applications of the fashion domain, natural language processing technology has become contributed in the area, so that it can build the connection between clothes image and human semantic understandings. An element work for combing images and language understanding is how to generate a natural language sentence that accurately summarizes the contents of a clothes image. In this paper, we develop a joint attribute detection and visual attention framework for clothes image captioning. Specifically, in order to involve more attributes of clothes to learn, we first utilize a pre-trained Convolutional Neural Network (CNN) to learn the feature that can characterize more information about clothing attribute. Based on such learned feature, we then adopt an encoder/decoder framework, where we first encoder the feature of clothes and then and input it to a language Long Short-Term Memory(LSTM) model for decoding the clothes descriptions. The method greatly enhances the performance of clothes image captioning and reduces the misleading attention. Extensive simulations based on real-world data verify the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

• Rectification of problems due to sudden illumination changes and proper alignment of feature set for an effective FER. • Isolation of foreground from the background of an image to facilitate the facial key points extraction. • Elimination of unwanted noise and smoothen by using AMF filtering technique. • MATP-based texture pattern and DESURF-LOP-based facial key points to the FER outperform the traditional methods. • PPCSO-based feature selection and LSTM-CNN based classification recognize the expressions accurately. Facial Expression Recognition (FER) plays the vital role in the Human Computer Interface (HCI) applications. The illumination and pose variations affect the FER adversely. The projection of complex 3D actions on the image plane and the inaccurate alignment are the major issues in the FER process. This paper presents the novel Multi-Angle Optimal Pattern-based Deep Learning (MAOP-DL) method to rectify the problem from sudden illumination changes, find the proper alignment of a feature set by using multi-angle-based optimal configurations. The proposed method includes the five major processes as Extended Boundary Background Subtraction (EBBS), Multi-Angle Texture Pattern+STM, Densely Extracted SURF+Local Occupancy Pattern (LOP), Priority Particle Cuckoo Search Optimization (PPCSO) and Long Short-Term Memory -Convolutional Neural Network (LSTM-CNN). Initially, the EBBS algorithm subtracts the background and isolates the foreground from the images which overcome the illumination and pose variation. Then, the MATP-STM extracts the texture patterns and DESURF-LOP extracts the relevant key features of the facial points. The PPCSO algorithm selects the relevant features from the MATP-STM feature set to speed up the classification. The employment of LSTM-CNN predicts the required label for the facial expressions.The major key findings of the proposed work are clear image analysis, effective handling of pose/illumination variations and the facial alignment. The proposed MAOP-DL validates its effectiveness on two standard databases such as CK+ and MMI regarding various metrics and confirm their assurance of wide applicability in recent applications. [ABSTRACT FROM AUTHOR]

• An effective method for detection of multi-class brain abnormalities is proposed. • The proposed deep CNN model demands less learnable parameters. • The CNN helps to learn prominent features automatically from the brain MRIs. • The model achieves greater classification accuracy over benchmark datasets. Automated detection of multi-class brain abnormalities through magnetic resonance imaging (MRI) has received much attention due to its clinical significance and therefore has become an active area of research over the years. The earlier automated schemes often followed traditional machine learning paradigms, in which the proper choice of features and classifiers has remained a major concern. Therefore, deep learning algorithms have been profoundly applied in various medical imaging applications. In this paper, a deep convolutional neural network (CNN) based automated approach is designed for the diagnosis of multi-class brain abnormalities. The proposed CNN model comprises five layers with learnable parameters: four convolutional layers and one fully-connected layer. The objective of designing such a custom deep network is to achieve greater classification performance with reduced number of parameters. The proposed model is evaluated on two benchmark multi-class brain MRI datasets namely, MD-1 and MD-2. The model achieved a classification accuracy of 100.00% and 97.50% on MD-1 and MD-2 datasets respectively. Moreover, four pre-trained CNN models based on the transfer learning approach have been tested over the same datasets. The comparative analysis with existing schemes indicates the superiority of the proposed method. [ABSTRACT FROM AUTHOR]

• A novel frame extraction mechanism that works on videos on the fly. • Extract frames from input video containing human actions only with no delay. • A typical CNN heuristic for efficient human action recognition. • Performance in various measures found to be outperforming among competent schemes. Recognition of human actions with optimum accuracy and with less computational overhead has always been a challenging task. A suitable framework towards this problem can provide a robust solution for numerous domains of expertise that demands automatic surveillance. In this paper, a two-fold framework has been proposed for the frame extraction and recognition of human action therein from a video input. A Fuzzy inferencing technique has been suitably used for the purpose of video frame extraction in a smarter way. Frames are extracted from a video only when an event of action is initiated and about to commit. By this the additional computational load of continuous frame extraction is eased efficiently. Subsequent to this process, we also introduce a typical convolutional neural network (CNN) that is used for the purpose of human action recognition. This typical CNN has been designed so as to suit the nature of input that is supplied to it. Experimental evaluation has been performed on the standard HMDB51 database. A total of eight distinct and prominent actions which are mostly common in daily routine of human life are considered for the purpose. Performance comparison of the proposed scheme with other state of the art schemes is also performed with three different measures. Evaluation results for the proposed scheme indicates overall rate of accuracies of 96.5% and 98% for frame extraction and action recognition respectively. The proposed scheme outperformed the others in most of the measuring aspects. [ABSTRACT FROM AUTHOR]

• Promising pruning methods utilize empirical classification loss to remove parameters of CNNs. • Empirical classification loss, however, might be vulnerable to outlier samples. • We propose a novel pruning criterion to be robust against the outlier samples. • Our method also provides stable metric for parameters. • Experiment results demonstrate favorable performance of our method in comparison with the other pruning methods. Deep convolutional neural network (CNN) with considerable number of parameters is one of the promising methods for image recognition. There, however, is generally difficult in applying deep CNNs to resource constrained devices due to the heavy computational burden. For reducing computational cost of CNNs while retaining the classification performance, it is effective to apply pruning methods that remove from CNNs redundant parameters less contributing to classification. The contribution of parameters can be estimated by the empirical classification loss computed over training samples to which ground-truth labels are assigned. The empirical classification loss, however, might be vulnerable to the outlier samples and/or the hard ones that are difficult to classify, and thus the pruning would accordingly be degraded. In this paper, we propose a pruning method based on a novel criterion to measure the redundancy of the parameters in CNNs through empirical classification loss. We start with the Taylor expansion of the loss function and then derive the mathematical formulation of the pruning criterion so as to be robust against some sort of outlier samples. The proposed pruning criterion can also provide stable metric for parameters and evaluate layers of various depth fairly without biases toward shallower or deeper layers. In addition, we present an effective method to normalize the criterion scores for further improving performance. In the experiments on image classification, our method exhibits favorable performance compared with the other methods. [ABSTRACT FROM AUTHOR]

• A novel polycrystalline silicon wafer defect segmentation method based on CNNs is proposed. • The proposed method alleviates unbalanced pixel distribution by adopting region proposal network. • The experimental results on real defect images show that the proposed methods outperforms state-of-the-art methods. Defect segmentation is an important way for defect detection in machine vision. For polycrystalline silicon wafer production, it is difficult to automatically segment defects due to its inhomogeneous background and unpredictable defect shapes. The conventional methods based on handcrafted models or features not only heavily rely on the expertise, but also are not flexible from one application case to another. In this paper, we propose a defect segmentation method for polycrystalline silicon wafer based on the deep convolutional networks. Firstly, we apply Region Proposal Network (RPN) to generate underlying defect regions. Then, these generated image patches are processed to suitable sizes for feeding into a improved segmentation network which is modified based on U-net and a dilation convolution. Real defect images are used to test the proposed method and experimental results show that proposed method achieves better performance compared with the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

• A joint spatial-spectral hyperspectral image classification framework based on convolutional neural network is proposed. • A two-stream convolutional network is designed to learn the spatial-spectral features at different scales. • A spatial enhancement strategy is proposed to obtain more comprehensive spatial features with limited training samples. Hyperspectral image (HSI) classification technology has been widely used in many earth observation tasks, such as detection, recognition, and surveillance. The traditional hyperspectral image classification methods mainly utilize hand-crafted features, such as edge and texture descriptors, which are not robust for different input data. By contrast, deep learning based methods exploit high-level features for hyperspectral image classification, but they usually degenerate the spatial-spectral structure, depend on a large number of training samples, and ignore a large amount of implicitly useful information. To address these problems, a new joint spatial-spectral hyperspectral image classification method based on different-scale two-stream convolutional network and spatial enhancement strategy is proposed in this paper. First, the pixel blocks at different scales around the center pixel are selected as the basic units to be processed. Then, a spatial enhancement strategy is designed to obtain various spatial location information under the limited training samples by the spatial rotation and row-column transformation. Finally, the spatial-spectral feature is learned by a different-scale two-stream convolutional network, and the classification result of the center pixel is obtained by a softmax layer. Experimental results on two datasets demonstrate that the proposed method outperforms other state-of-the-art methods qualitatively and quantitatively. [ABSTRACT FROM AUTHOR]

• A framework for automatic organization and classification of educational videos. • Collection of a large-scale MOOC dataset. • Thorough investigation of data representation techniques for video lecture classification. • Performance evaluation of deep and conventional classifiers on MOOC dataset. The advent of MOOC platforms brought an abundance of video educational content that made the selection of best fitting content for a specific topic a lengthy process. To tackle this challenge in this paper we report our research efforts of using deep learning techniques for managing and classifying educational content for various search and retrieval applications in order to provide a more personalized learning experience. In this regard, we propose a framework which takes advantages of feature representations and deep learning for classifying video lectures in a MOOC setting to aid effective search and retrieval. The framework consists of three main modules. The first module called pre-processing concerns with video-to-text conversion. The second module is transcript representation which represents text in lecture transcripts into vector space by exploiting different representation techniques including bag-of-words, embeddings, transfer learning, and topic modeling. The final module covers classifiers whose aim is to label video lectures into the appropriate categories. Two deep learning models, namely feed-forward deep neural network (DNN) and convolutional neural network (CNN) are examined as part of the classifier module. Multiple simulations are carried out on a large-scale real dataset using various feature representations and classification techniques to test and validate the proposed framework. [ABSTRACT FROM AUTHOR]

• This paper proposes a fully convolutional deep architecture for predicting depth and surface normal from a single RGB image. • A state-of-the-art CNN, DenseNet, is employed to extract the deep convolutional features from the input RGB images. • Deconvolution is applied at various stages of the network to enhance the spatial resolution of intermediate feature maps. • Bottleneck layers are incorporated to acquire some feature maps of same resolution as that of the deconvolution blocks. • Concatenated features are progressed along the deep network in a cascaded manner to yield the target (depth, surface normal). Understanding the 3D perspective of a scene is imperative in improving the precision of intelligent autonomous systems. The difficulty in understanding is compounded when only one image of the scene is available at disposal. In this regard, we propose a fully convolutional deep framework for predicting the depth map and surface normal from a single RGB image in a common architecture. The proposed model accepts an input RGB image of size 320 × 240 × 3, and predicts the output via 4 stages: (1) It forwards the input through the convolution layers of DenseNet-161 to extract deep features of size 10 × 7 × 2208, (2) Few deconvolution blocks are appended at the end of the DenseNet in order to enhance the resolution of target output, (3) The output feature maps of the bottleneck layers, which are appended to the dense blocks of the DenseNet to facilitate feature-reuse, are concatenated with same resolution feature maps of deconvolution blocks, (4) These concatenated features are progressed along the network in a cascaded manner to form the final output of resolution 160 × 120. Understanding the 3D perspective of a scene is imperative in improving the precision of intelligent autonomous systems. The difficulty in understanding is compounded when only one image of the scene is available at disposal. In this regard, we propose a fully convolutional deep framework for predicting the depth map and surface normal from a single RGB image in a common architecture. The DenseNet CNN architecture is employed to learn the complex mapping between an input RGB image and its corresponding 3D primitives. We introduce a novel approach of multi-stage cascaded deconvolution, where the output feature maps of one dense block are reused by concatenating with the feature maps of the corresponding deconvolution block. These combined feature maps are progressed along the deep network in a pre-activated manner to construct the final output. The network is trained separately for estimating depth and surface normal while keeping the architecture same. The suggested architecture, compared to the counterparts, uses fewer training samples and model parameters. Exhaustive experiments on benchmark dataset not only reveal the efficacy of the proposed multi-stage scheme over the one-way sequential deconvolution but also outperform the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

The use of Convolutional Neural Networks (CNNs) in image analysis and recognition paved the way for long-term visual place recognition. The transferable power of generic descriptors extracted at different layers of off-the-shelf CNNs has been successfully exploited in various visual place recognition scenarios. In this paper we tackle this problem by extracting the full output of an intermediate layer and building an image descriptor of lower dimensionality by omitting the activation of filters corresponding to environmental changes. Thus, we are able to increase the robustness of the cross-season visual place recognition. We test our approach on the Nordland dataset, the biggest and the most challenging dataset up to date, where the included four seasons induce great illumination and appearance changes. The experiments show that our new approach can significantly increase, up to 14%, the baseline (single-image search) performance of deep features. [ABSTRACT FROM AUTHOR]

This paper presents a system for human posture recognition using a camera and two infrared light sources. It uses as input the combination of the body silhouette and its (invisible to the eye) cast shadows. Conventional video-surveillance methods based on a single camera can fail to infer the correct posture since different postures can look similar under perspective projection. Fortunately, cast body shadows, generated by infrared lights, offer additional posture information that cannot be directly captured by a single camera. Each shadow can be projected on different surfaces (e.g. floor, walls, and furniture) generating complex body projections that represent various shapes within the same posture class. These shadow images are very challenging and difficult to describe with traditional handcrafted features that need to be somewhat invariant to these within-class changes. However, a deep convolution neural network (CNN) is able to learn a better data representation from a large-scale dataset. In the absence of a big real dataset, we propose to use synthetic data for training the CNN classifier. Learning from synthetic data is a challenging task due to the gap between synthetic and real feature distributions. Thus, we propose a normalization technique to bridge the gap and help the classifier to better generalize with real data. We evaluated the proposed system on a new real dataset captured in our laboratory and a simulated dataset generated with computer graphics tools. Experimental results validated the efficiency of the CNN model against other conventional methods. Furthermore, the combination of cast shadows and body silhouette had better performance than using only the body silhouette as expected. [ABSTRACT FROM AUTHOR]

Scene character recognition remains a challenging task due to many interference factors. Considering that characters are composed of a series of parts arranged in certain structures, in this paper, we propose a novel representation termed multi-order co-occurrence activations (MCA) encoded with Fisher Vector (FV), namely MCA-FV. It implicitly models the co-occurrence information of discriminative character parts at different orders to boost the recognition performance. We first extract convolutional activations as local descriptors of character parts from convolutional neural networks (CNNs). Then, we introduce MCA features to capture the multi-order co-occurrence cues among different discriminative character parts. Finally, we apply FV to encode co-occurrence features of each order and obtain a global representation of MCA-FV. The proposed method is evaluated on four scene character datasets including English and Chinese datasets. Experiment results demonstrate the effectiveness of the proposed method for scene character recognition. [ABSTRACT FROM AUTHOR]

Data fusion from different modalities has been extensively studied for a better understanding of multimedia contents. On one hand, the emergence of new devices and decreasing storage costs cause growing amounts of data being collected. Though bigger data makes it easier to mine information, methods for big data analytics are not well investigated. On the other hand, new machine learning techniques, such as deep learning, have been shown to be one of the key elements in achieving state-of-the-art inference performances in a variety of applications. Therefore, some of the old questions in data fusion are in need to be addressed again for these new changes. These questions are: What is the most effective way to combine data for various modalities? Does the fusion method affect the performance with different classifiers? To answer these questions, in this paper, we present a comparative study for evaluating early and late fusion schemes with several types of SVM and deep learning classifiers on two challenging RGB-D based visual recognition tasks: hand gesture recognition and generic object recognition. The findings from this study provide useful policy and practical guidance for the development of visual recognition systems. [ABSTRACT FROM AUTHOR]

The paradigm of Convolutional Neural Network (CNN) has already shown its potential for many challenging applications of computer vision, such as image classification, object detection and action recognition. In this paper, the task of 3D model retrieval is addressed by exploiting such promising paradigm. However, 3D models are usually represented with a collection of orderless points, lines and surfaces in a three dimensional space, which makes it difficult to involve the operation of convolution, pooling, etc . Yet, we propose a practical and effective way for applying CNN to 3D model retrieval, by training the network with the depth projections of 3D model. This CNN is regarded as a generic feature extractor for depth image. With large amounts of training data, the learned feature, which is called Neural Shape Codes, can handle various deformation changes that exist in shape analysis. The reported experimental results on several 3D shape benchmark datasets show the superior performance of the proposed method. [ABSTRACT FROM AUTHOR]