Search

Music emotion recognition is becoming an important research direction due to its great significance for music information retrieval, music recommendation, and so on. In the task of music emotion recognition, the key to achieving accurate emotion recognition lies in how to extract the affect-salient features fully. In this paper, we propose an end-to-end spatial-temporal feature extraction method named FFA-BiGRU for music emotion classification. Taking the log Mel-spectrogram of music audio as the input, this method employs an attention-based convolutional residual module named FFA, which serves as a spatial feature learning module to obtain multi-scale spatial features. In the FFA module, three group architecture blocks extract multi-level spatial features, each of which consists of a stack of multiple channel-spatial attention-based residual blocks. Then, the output features from FFA are fed into the bidirectional gated recurrent units (BiGRU) module to capture the temporal features of music further. In order to make full use of the extracted spatial and temporal features, the output feature maps of FFA and those of the BiGRU are concatenated in the channel dimension. Finally, the concatenated features are passed through fully connected layers to predict the emotion classification results. The experimental results of the EMOPIA dataset show that the proposed model achieves better classification accuracy than the existing baselines. Meanwhile, the ablation experiments also demonstrate the effectiveness of each part of the proposed method. [ABSTRACT FROM AUTHOR]

Due to the widespread applications of change detection technology in urban change analysis, environmental monitoring, agricultural surveillance, disaster detection, and other domains, the task of change detection has become one of the primary applications of Earth orbit satellite remote sensing data. However, the analysis of dual-temporal change detection (CD) remains a challenge in high-resolution optical remote sensing images due to the complexities in remote sensing images, such as intricate textures, seasonal variations in imaging time, climatic differences, and significant differences in the sizes of various objects. In this paper, we propose a novel U-shaped architecture for change detection. In the encoding stage, a multi-branch feature extraction module is employed by combining CNN and transformer networks to enhance the network's perception capability for objects of varying sizes. Furthermore, a multi-branch aggregation module is utilized to aggregate features from different branches, providing the network with global attention while preserving detailed information. For dual-temporal features, we introduce a spatiotemporal discrepancy perception module to model the context of dual-temporal images. Particularly noteworthy is the construction of channel attention and token attention modules based on the transformer attention mechanism to facilitate information interaction between multi-level features, thereby enhancing the network's contextual awareness. The effectiveness of the proposed network is validated on three public datasets, demonstrating its superior performance over other state-of-the-art methods through qualitative and quantitative experiments. [ABSTRACT FROM AUTHOR]

Remote sensing image classification plays a crucial role in the field of remote sensing interpretation. With the exponential growth of multi-source remote sensing data, accurately extracting target features and comprehending target attributes from complex images significantly impacts classification accuracy. To address these challenges, we propose a Canny edge-enhanced multi-level attention feature fusion network (CAF) for remote sensing image classification. The original image is specifically inputted into a convolutional network for the extraction of global features, while increasing the depth of the convolutional layer facilitates feature extraction at various levels. Additionally, to emphasize detailed target features, we employ the Canny operator for edge information extraction and utilize a convolution layer to capture deep edge features. Finally, by leveraging the Attentional Feature Fusion (AFF) network, we fuse global and detailed features to obtain more discriminative representations for scene classification tasks. The performance of our proposed method (CAF) is evaluated through experiments conducted across three openly accessible datasets for classifying scenes in remote sensing images: NWPU-RESISC45, UCM, and MSTAR. The experimental findings indicate that our approach based on incorporating edge detail information outperforms methods relying solely on global feature-based classifications. [ABSTRACT FROM AUTHOR]

Hazy images often cause blurring, detail loss and color distortion, which makes it difficult to address the other visual tasks such as tracking, classification and object detection. In recent years, significant advances have been made in image dehazing task, dominated by convolutional neural networks (CNNs). Most existing CNNs methods tend to estimate the transmission map and atmospheric light and then recover the haze-free image based on atmospheric scattering model. However, its dehazing performance is limited due to inaccurate estimation. To this end, we present a new architecture called multi-level features and adaptive fusion network (MFAF-Net) for single image dehazing, which can obtain the haze-free image in an end-to-end manner. For one thing, we utilize a novel context enhanced module as the core of feature extraction, and it combines multi-scale dilation convolution layers with feature attention module, which enables to acquire much informative contextual information. For another, we present a new fusion approach called adaptive fusion module for both low- and high-level feature fusion, and it can provide more flexibility when handling features of inconsistent semantic and level; thus, our network restores images with more detailed information. Experimental results on both synthetic and real-world datasets demonstrate that MFAF-Net outperforms existing state-of-the-art methods in terms of quantitative and qualitative evaluation metrics. The code will be made publicly available on Github. [ABSTRACT FROM AUTHOR]

Music emotion recognition is becoming an important research direction due to its great significance for music information retrieval, music recommendation, and so on. In the task of music emotion recognition, the key to achieving accurate emotion recognition lies in how to extract the affect-salient features fully. In this paper, we propose an end-to-end spatial-temporal feature extraction method named FFA-BiGRU for music emotion classification. Taking the log Mel-spectrogram of music audio as the input, this method employs an attention-based convolutional residual module named FFA, which serves as a spatial feature learning module to obtain multi-scale spatial features. In the FFA module, three group architecture blocks extract multi-level spatial features, each of which consists of a stack of multiple channel-spatial attention-based residual blocks. Then, the output features from FFA are fed into the bidirectional gated recurrent units (BiGRU) module to capture the temporal features of music further. In order to make full use of the extracted spatial and temporal features, the output feature maps of FFA and those of the BiGRU are concatenated in the channel dimension. Finally, the concatenated features are passed through fully connected layers to predict the emotion classification results. The experimental results of the EMOPIA dataset show that the proposed model achieves better classification accuracy than the existing baselines. Meanwhile, the ablation experiments also demonstrate the effectiveness of each part of the proposed method.

Due to the widespread applications of change detection technology in urban change analysis, environmental monitoring, agricultural surveillance, disaster detection, and other domains, the task of change detection has become one of the primary applications of Earth orbit satellite remote sensing data. However, the analysis of dual-temporal change detection (CD) remains a challenge in high-resolution optical remote sensing images due to the complexities in remote sensing images, such as intricate textures, seasonal variations in imaging time, climatic differences, and significant differences in the sizes of various objects. In this paper, we propose a novel U-shaped architecture for change detection. In the encoding stage, a multi-branch feature extraction module is employed by combining CNN and transformer networks to enhance the network’s perception capability for objects of varying sizes. Furthermore, a multi-branch aggregation module is utilized to aggregate features from different branches, providing the network with global attention while preserving detailed information. For dual-temporal features, we introduce a spatiotemporal discrepancy perception module to model the context of dual-temporal images. Particularly noteworthy is the construction of channel attention and token attention modules based on the transformer attention mechanism to facilitate information interaction between multi-level features, thereby enhancing the network’s contextual awareness. The effectiveness of the proposed network is validated on three public datasets, demonstrating its superior performance over other state-of-the-art methods through qualitative and quantitative experiments.

Remote sensing image classification plays a crucial role in the field of remote sensing interpretation. With the exponential growth of multi-source remote sensing data, accurately extracting target features and comprehending target attributes from complex images significantly impacts classification accuracy. To address these challenges, we propose a Canny edge-enhanced multi-level attention feature fusion network (CAF) for remote sensing image classification. The original image is specifically inputted into a convolutional network for the extraction of global features, while increasing the depth of the convolutional layer facilitates feature extraction at various levels. Additionally, to emphasize detailed target features, we employ the Canny operator for edge information extraction and utilize a convolution layer to capture deep edge features. Finally, by leveraging the Attentional Feature Fusion (AFF) network, we fuse global and detailed features to obtain more discriminative representations for scene classification tasks. The performance of our proposed method (CAF) is evaluated through experiments conducted across three openly accessible datasets for classifying scenes in remote sensing images: NWPU-RESISC45, UCM, and MSTAR. The experimental findings indicate that our approach based on incorporating edge detail information outperforms methods relying solely on global feature-based classifications.

Abstract Fully convolutional neural networks have achieved great success in salient object detection, in which the effective use of multi‐layer features plays a critical role. Based on this advantage, many saliency detectors have emerged in recent years, and most of them designed a series of network structures to integrate the multi‐level features generated by the backbone network. However, information in different layer play different roles in saliency object detection, how to integrate them effectively is still a great challenge. In this article, a selective feature fusion network which consists of a selective feature fusion module (SFM) and an attention‐guide hierarchical feature emphasis module (AEM) is proposed. Most of the previous works mainly integrate multi‐level feature by addition and concatenation, as a difference, SFM adaptively selects the important information from the input features in the fusion, which effectively avoids introducing too much redundant information. Besides, AEM combines spatial attention and channel attention to enhance features simply and effectively by hierarchical iteration, and further improve the accuracy of salient object detection. Experiments on five datasets show that the proposed selective feature fusion method achieve satisfactory results when comparing to other state‐of‐the‐art salient object detection approaches.

Copyright of Journal of Computer Engineering & Applications is the property of Beijing Journal of Computer Engineering & Applications Journal Co Ltd. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Fully convolutional neural networks have achieved great success in salient object detection, in which the effective use of multi‐layer features plays a critical role. Based on this advantage, many saliency detectors have emerged in recent years, and most of them designed a series of network structures to integrate the multi‐level features generated by the backbone network. However, information in different layer play different roles in saliency object detection, how to integrate them effectively is still a great challenge. In this article, a selective feature fusion network which consists of a selective feature fusion module (SFM) and an attention‐guide hierarchical feature emphasis module (AEM) is proposed. Most of the previous works mainly integrate multi‐level feature by addition and concatenation, as a difference, SFM adaptively selects the important information from the input features in the fusion, which effectively avoids introducing too much redundant information. Besides, AEM combines spatial attention and channel attention to enhance features simply and effectively by hierarchical iteration, and further improve the accuracy of salient object detection. Experiments on five datasets show that the proposed selective feature fusion method achieve satisfactory results when comparing to other state‐of‐the‐art salient object detection approaches. [ABSTRACT FROM AUTHOR]

Purpose: Computed tomography coronary angiography (CCTA) images provide optimal visualization of coronary arteries to aid in diagnosing coronary heart disease (CHD). With the deep convolutional neural network, this work aims to develop an intelligent and lightweight coronary artery segmentation algorithm that can be deployed in hospital systems to assist clinicians in quantitatively analyzing CHD. Methods: With the multi-level feature fusion, we proposed Dual-Attention Coordination U-Net (DAC-UNet) that achieves automated coronary artery segmentation in 2D CCTA images. The coronary artery occupies a small region, and the foreground and background are extremely unbalanced. For this reason, the more original information can be retained by fusing related features between adjacent layers, which is conducive to recovering the small coronary artery area. The dual-attention coordination mechanism can select valid information and filter redundant information. Moreover, the complementation and coordination of double attention factors can enhance the integrity of features of coronary arteries, reduce the interference of non-coronary arteries, and prevent over-learning. With gradual learning, the balanced character of double attention factors promotes the generalization ability of the model to enhance coronary artery localization and contour detail segmentation. Results: Compared with existing related segmentation methods, our method achieves a certain degree of improvement in 2D CCTA images for the segmentation accuracy of coronary arteries with a mean Dice index of 0.7920. Furthermore, the method can obtain relatively accurate results even in a small sample dataset and is easy to implement and deploy, which is promising. The code is available at: https://github.com/windfly666/Segmentation. Conclusion: Our method can capture the coronary artery structure end-to-end, which can be used as a fundamental means for automatic detection of coronary artery stenosis, blood flow reserve fraction analysis, and assisting clinicians in diagnosing CHD. [ABSTRACT FROM AUTHOR]

Light field (LF) images suffer from low spatial resolution due to the trade-off between angular and spatial resolutions. Thus, spatial super-resolution (SR) of LF images is an essential task to obtain high-quality LF images. However, the existing SR networks still have limitations, since they exploit only single-level features to use sub-pixel information in LF images. In this paper, we proposed a light field super-resolution (LFSR) network to effectively improve the spatial resolution of light field images. The proposed network takes one target image and its 8-neighboring images for references. We construct multi-level structures for the proposed network to effectively estimate and mix sub-pixel information in reference images. The proposed network is composed of a feature extractor, a feature warping module, a feature mixing module, and a upscaling module. The feature extractor provides multi-level features for SR and offsets to the feature warping module to obtain aligned features for multiple reference images. The feature mixing module mixes multiple aligned features based on the similarity between the target and reference images to obtain multi-level mixed features. Finally, the upscaling module generates a high-resolution residual image using the multi-level mixed features. Experimental results demonstrate the proposed network outperforms the state-of-the-art methods on various light field datasets. The pre-trained model and source codes are available at https://github.com/Hwa-Jong/LF_MLS.

Training a model that can quickly adapt to new tasks keeps a crucial challenge for few-shot learning. Approaches based on metric-learning are very popular and promising. However, the existing such approaches usually only rely on the feature information obtained in the last layer of the feature extraction backbone for similarity metric, and do not consider that the feature information obtained in multi-layered can be fully utilized. Furthermore, the targeted differentiation of features and the selection and construction of loss functions are usually ignored by these approaches, which will become important factors that limit the performance of the model. Therefore, a few-shot learning approach MFR-MFC with attention mechanism based on multi-level feature representation and multi-layered fusion contrast is proposed in this paper. First, multi-level feature representations are introduced when extracting features, feature information of multi-layers are used to perform information fusion, and then this information is utilized for subsequent metric. Then, when training the model, multi-level features are also used to introduce multi-layered fusion contrast. Additionally, an attention module is introduced in the feature extraction process to make the model obtain more discriminative information. Experiments have shown that the approach proposed in this paper has achieved excellent performance in few-shot classification and has significant advantages compared with advanced technologies. [ABSTRACT FROM AUTHOR]

The performance of the salient object detection of strip surface defects has been promoted largely by deep learning based models. However, due to the complexity of strip surface defects, the existing models perform poorly in the challenging scenes such as noise disturbance, and low contrast between defect regions and background. Meanwhile, the detection results of existing models often suffer from coarse boundary details. Therefore, we propose a novel saliency model, namely an Edge-aware Multi-level Interactive Network, to detect the defects from the strip steel surface. Concretely, our model adopts the U-shape architecture where the two crucial points are the interactive feature integration and the edge-guided saliency fusion. Firstly, except the skip connection that combines the same stage of encoder and decoder, we deploy another connection, where the features from adjacent levels of encoder are transferred to the same stage of decoder. By this way, we are able to provide an effective fusion of multi-level deep features, yielding a well depiction for defects. Secondly, to give well-defined boundaries for prediction results, we add the edge extraction branch after each decoder block, where the progressive feature aggregation endows the edge with precise details and complete object cues. Meanwhile, together with the edge extraction branches, we deploy the saliency prediction branch at each decoder stage. After that, coupled with the fine edge information, we fuse all outputs of saliency prediction branches into the final saliency map, where the edge cue steers the saliency result to pay more attention to the boundary details. Following this way, we can provide a high-quality saliency map which can accurately locate and segment the defects. Extensive experiments are performed on the public dataset, and the results prove the effectiveness and robustness of our model which consistently outperforms the state-of-the-art models.

High-resolution is essential to achieve remarkable performance in many applications using hyperspectral images. However, the acquired hyperspectral images are often low-resolution due to the limitations of sensors. Recently, deep learning-based methods have been widely studied to address the super-resolution problem, and most super-resolution methods are learned with the simulated dataset. However, the pre-defined down-sampling function employed in the simulation is simple, thus the trained model often suffers from the poor generalization for real-world images. In this paper, we propose an unsupervised super-resolution approach that does not require the paired dataset. The method is conducted in two stages: unsupervised image generation and supervised image super-resolution. Specifically, the image generation model generates a low-resolution image through a generative adversarial network in an unsupervised manner. Then, the generated low-resolution images are used to train the super-resolution model with high-resolution images in a supervised manner. To explore the high correlation of hyperspectral image in spectral and spatial domain, group convolution, attention mechanism, and multi-level features are employed in the super-resolution model. Experiments on the public dataset CAVE and Harvard show that the proposed model provides outperforming super-resolution ability over the compared methods. Also, the results on images with unknown degradation show a promising generalization of the proposed model. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Zhengzhou University: Engineering Science is the property of Editorial Office of Journal of Zhengzhou University: Engineering Science and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Person re-identification has become a challenging task due to various factors. One key to effective person re-identification is the extraction of the discriminative features of a person's appearance. Most previous works based on deep learning extract pedestrian characteristics from neural networks but only from the top feature layer. However, the low-layer feature could be more discriminative in certain circumstances. Hence, we propose a method, named the multi-level feature network with multiple losses (MFML), which has a multi-branch network architecture that consists of multiple middle layers and one top layer for feature representations. To extract the discriminative middle-layer features and have a good effect on deeper layers, we utilize the triplet loss function to train the middle-layer features. For the top layer, we focus on learning more discriminative feature representations, so we utilize the hybrid loss (HL) function to train the top-layer feature. Instead of concatenating multilayer features directly, we concatenate the weighted middle-layer features and the weighted top-layer feature as the discriminative features in the testing phase. The extensive evaluations conducted on three datasets show that our method achieves a competitive accuracy level compared with the state-of-the-art methods.

Magnetic resonance imaging (MRI) is an essential radiology technique in clinical diagnosis, but its spatial resolution may not suffice to meet the growing need for precise diagnosis due to hardware limitations and thicker slice thickness. Therefore, it is crucial to explore suitable methods to increase the resolution of MRI images. Recently, deep learning has yielded many impressive results in MRI image super-resolution (SR) reconstruction. However, current SR networks mainly use convolutions to extract relatively single image features, which may not be optimal for further enhancing the quality of image reconstruction. In this work, we propose a multi-level feature extraction and reconstruction (MFER) method to restore the degraded high-resolution details of MRI images. Specifically, to comprehensively extract different types of features, we design the triple-mixed convolution by leveraging the strengths and uniqueness of different filter operations. For the features of each level, we then apply deconvolutions to upsample them separately at the tail of the network, followed by the feature calibration of spatial and channel attention. Besides, we also use a soft cross-scale residual operation to improve the effectiveness of parameter optimization. Experiments on lesion-free and glioma datasets indicate that our method obtains superior quantitative performance and visual effects when compared with state-of-the-art MRI image SR methods., Competing Interests: Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper., (Copyright © 2024 Elsevier Ltd. All rights reserved.)

Copyright of Journal of Computer-Aided Design & Computer Graphics / Jisuanji Fuzhu Sheji Yu Tuxingxue Xuebao is the property of Gai Kan Bian Wei Hui and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

This study proposes a method to recognize façades from large-scale urban scenes based on multi-level image features utilizing a recently developed oblique aerial photogrammetry technique. The method involves the use of multi-level image features, a bottom-up feature extraction procedure to produce regions of interest through monoscopic analysis, and then a coarse-to-fine feature matching strategy to characterise and match the regions in a stereoscopic model. Feature extraction from typical urban Manhattan scenes is based on line segments. Windows are re-organised based on the spatial constraints of line segments and the homogeneous structure of the spectrum. Façades as regions of interest are successfully constructed with a remarkable single edge and evidence from windows to get rid of occlusion. Feature matching is hierarchically performed beginning from distinctive facades and regularly distributed windows to the sub-pixel point primitives. The proposed strategy can effectively solve ambiguity and multi-solution problems in the complex urban scene matching process, particularly repetitive and poor-texture façades in oblique view.

Electrical discharge machining (EDM) is widely used in mold manufacturing to form the intricate geometric features which are difficult to be produced by the conventional machining process. Owing to the use of a large number of electrodes, the cost and time consumption for electrode design and NC programming has become a bottleneck in boosting productivity of mold enterprises. Although computer-aided design (CAD)/computer-aided manufacturing (CAM) systems are widely used, because of the complex shapes and diversity of electrodes, manual operations are commonly employed to generate process plans and tool paths for electrode manufacturing, which is cumbersome and error-prone. In order to realize the intelligent manufacturing of electrodes in such a knowledge-intensive domain, this paper presents a feature-based integration approach of CAD/computer-aided process planning (CAPP)/CAM. A hierarchical taxonomy of electrode features is introduced and a hybrid feature recognition method is proposed to build multi-level feature tree. Then, feature knowledge and domain know-how based process planning and optimization are achieved followed by automatic tool path generation which aim to machining the electrode precisely and economically. In the framework of system integration, a structured product model is established to capture and encapsulate geometric entities, machining features, technical information, process plans, and measurement data of machining error to realize seamless flow of information among CAD/CAPP/CAM systems. The effectiveness and efficiency of the proposed approach are demonstrated by case studies and industry implementation. [ABSTRACT FROM AUTHOR]

The change detection task plays an irreplaceable role in the remote sensing field. However, most methods ignore the edge distinctive information. Such information is not only significant in some change detection tasks such as channel and river changes, but also important for refining the accuracy of change detection. Therefore, an edge refinement multi-feature (ERMF) extraction method, employing a siamese network to extract the primary discriminative features at five scales of bitemporal remote sensing images, is proposed in this paper. On the one hand, an edge refinement module is designed to obtain the edge change map as well as the final accurate region change map. On the other hand, a multi-level feature extraction module is introduced to acquire a coarse change map consisting of low-level location information and high-level semantic information at five different scales. Besides, it is worth emphasizing that we present a hybrid loss to evaluate the ERMF model. Experiments demonstrate that the ERMF model outperforms seven state-of-the-art methods in both qualitative and quantitative evaluations. • In this paper, a novel end-to-end remote sensing change detection model ERMF is proposed, which extracts edge change information and region change information. The results of regional change based on edge change information improved in the objective index, and boundary division is obvious in subjective evaluation. • Aiming at getting edge change information maps and refining the final results, an edge refinement module (ERM) is introduced. • In addition, a decent hybrid loss function is designed to constrain the results of edge change detection and region change detection. [ABSTRACT FROM AUTHOR]

Functional neuroimaging studies have found intra-regional activity and inter-regional connectivity alterations in patients with post-traumatic stress disorder (PTSD). However, the results of these studies are based on group-level statistics and therefore it is unclear whether PTSD can be discriminated at single-subject level, for instance using the machine learning approach. Here, we proposed a novel framework to identify PTSD using multi-level measures derived from resting-state functional MRI (fMRI). Specifically, three levels of measures were extracted as classification features: (1) regional amplitude of low-frequency fluctuations (univariate feature), which represents local spontaneous synchronous neural activity; (2) temporal functional connectivity (bivariate feature), which represents the extent of similarity of local activity between two regions, and (3) spatial functional connectivity (multivariate feature), which represents the extent of similarity of temporal correlation maps between two regions. Our method was evaluated on 20 PTSD patients and 20 demographically matched healthy controls. The experimental results showed that the features of each level could successfully discriminate PTSD patients from healthy controls. Furthermore, the combination of multi-level features using multi-kernel learning can further improve the classification performance. Specifically, the classification accuracy obtained by the proposed framework was 92.5 %, which was an increase of at least 5 and 17.5 % from the two-level and single-level feature based methods, respectively. Particularly, the limbic structure and prefrontal cortex provided the most discriminant features for classification, consistent with results reported in previous studies. Together, this study demonstrated for the first time that patients with PTSD can be identified at the individual level using resting-state fMRI data. The promising classification results indicated that this method may provide a complementary approach for improving the clinical diagnosis of PTSD. [ABSTRACT FROM AUTHOR]

Abstract: A dialog system is an intelligent program that helps users easily access information stored in a knowledge base by formulating requests in their natural language. A dialog system needs an intention prediction module for use as a preprocessor to reduce the search space of an automatic speech recognizer. To satisfy these needs, we propose a statistical model to predict speakers’ intentions. The proposed model represents a dialog history, with various levels of linguistic features. The proposed model predicts the user’s next intention by giving the linguistic features as inputs to a statistical machine learning model. In experiments conducted in a schedule management domain, the proposed model showed a higher average precision than the previous model. [Copyright &y& Elsevier]

Convolutional neural networks have shown outstanding effectiveness in person re-identification (re-ID). However, the models always have large number of parameters and much computation for mobile application. In order to relieve this problem, we propose a novel grafted network (GraftedNet), which is designed by grafting a high-accuracy rootstock and a light-weighted scion. The rootstock is based on the former parts of ResNet-50 to provide a strong baseline, while the scion is a new designed module, composed of the latter parts of SqueezeNet, to compress the parameters. To extract more discriminative feature representation, a joint multi-level and part-based feature is proposed. In addition, to train GraftedNet efficiently, we propose an accompanying learning method, by adding an accompanying branch to train the model in training and removing it in testing for saving parameters and computation. On three public person re-ID benchmarks (Market1501, DukeMTMC-reID and CUHK03), the effectiveness of GraftedNet is evaluated and its components are analyzed. Experimental results show that the proposed GraftedNet achieves 93.02%, 85.3% and 76.2% in Rank-1 and 81.6%, 74.7% and 71.6% in mAP, with only 4.6M parameters. • GraftedNet grafts a high-accuracy network and a light-weighted network. • Joint multi-level and part-based feature can improve person feature representation. • Accompanying branch helps GraftedNet to train better; can be removed in testing. • GraftedNet has better performance than original ResNet-50 with 4.6M parameters. [ABSTRACT FROM AUTHOR]