| Person re-identification(Re-ID)is a technology that solves the problem of iden-tification and retrieval of pedestrians across cameras and scenes.This technology has recently received more and more attention of researchers in the field of computer vision.There are two key problems in person re-identification:one is how to obtain a robust feature representation of person images;the other is how to design a good distance metric to evaluate the similarity between pedestrian images.The traditional Re-ID al-gorithms mainly perform the Re-ID task based on the handcraft feature descriptions and distance metrics.In recent years,convolutional neural networks have been increasingly applied to Re-ID systems due to their powerful capability of feature representation.A lot of recent works have shown that the performance of Re-ID task can be sig-nificantly improved by using the convolutional neural network model to simultaneously learn and fuse the local and global feature representations of pedestrian images.How-ever,existing local and global feature fusion models not only complicate the network designment,but also increase the dimension of feature expression of the person im-age after feature fusion.Moreover,existing feature matching methods only measure the similarity between the corresponding coordinate regions in two pedestrian images,ignoring the fact that the occlusions and the incorrect detections of persons,which in-creases the difficulties in evaluating the similarity of the two images.The computational complexity of existing Re-ID models with promising performance is too high to meet the requirements of processing large-scale Re-ID data in real time.Therefore,in order to meet the real-time requirements of Re-ID systems,an effective acceleration algorithm is demanded.In this paper,three different algorithms are proposed to alleviate the problems in existing Re-ID methods.They are Re-ID method based on local convolutional neural network,Re-ID method based on saliency region detection and matching,and a quanti-zation method based on nonlinear function of neural network model.The first method aims to solve the problems in existing local and global feature fusion models.This method proposes a new local network building block to extract the local features of pedestrian images with only a small amount of prior information.And,the fusion of local features and global features of the local building block can be completed on differ-ent hierarchical features without increasing the dimension of feature expression of the image.The second method aims at solving the problem of unreasonable matching of pedestrian features in the retrieval stage of Re-ID.The essence of this method is to find salient regions in different parts of person bodies,and then use salient region informa-tion to conduct feature alignment in feature matching stage.This operation can solve the problem of unreasonable feature matching caused by occlusions or inaccurate per-son detections.In addition to improving the performance of Re-ID models,this paper also studies how to accelerate the neural network model to improve the real-time perfor-mance of the Re-ID model in practical applications.The third method aims to solve the problem of gradient mismatch and the inability to quantize activation in existing model acceleration algorithms.This method explains neural network quantization from a new perspective and designs a differentiable non-linear quantization function for the quan-tization of neural network models.The training of model quantification is performed using a continuous relaxation method.In this paper,we compare the performance of these three proposed methods with existing algorithms through extensive experiments,and the results verify the effectiveness and advancement of our methods.The main contributions and novelties of this paper are summarized as follows:·Compared with Google Net,ResNet and DenseNet,we have designed a new neu-ral network building block that combines local and global information.This block can be easily and efficiently inserted in anywhere of any neural network model.The local convolution neural network model based on this local building block has achieved promising performance on the existing Re-ID datasets.·Compared with existing Re-ID models,this paper designed a feature extraction and matching method for human saliency region from the perspective of feature alignment.This method can solve the problem of model performance degradation caused by the pedestrian body absences or occlusions in person images in Re-ID tasks.·Compared with the existing low-bit quantization methods,we describe quantiza-tion for the first time as a differentiable nonlinear mapping function,and provide a simple/straightforward and universal solution for arbitrary bit weight and acti-vation quantization.This method does not encounter serious gradient mismatch problems,and can be learned in a lossless end-to-end manner. |
