Research On Person Re-identification Technology Based On Deep Learning

Computer Vision is an important branch of Artificial Intelligence.Its importance is selfevident.There are countless related studies around it.Based on the importance of human beings on self-related research,related research on human images in visual tasks is undoubtedly pivotal.Face Recognition is one of the earliest sub-topics in the research of visual tasks.Its research results have reached a very high level.However,it often has great limitations in real-world applications.Due to the many difficulties(such as shooting facial images are blurred,or there are large-scale occlusions in the face,etc.),it is not realistic to recognize a complete person only by the face.Under such circumstances,Person Re-Identification,as an important supplementary technology,can well make up for the shortcomings in the real-world applications of Face Recognition.Person Re-identification technology has an irreplaceable role in the construction of smart security,smart shopping malls,etc.As a major research hotspot of visual tasks,more and more researchers have begun to pay attention and research on it.In recent years,Deep Learning has developed rapidly,and the research on Person Reidentification task based on it has achieved excellent results.Compared with the traditional method,which is time-consuming and laborious to design features manually,deep learning method can mine person information and extract features based on a larger amount of data.The features extracted through the constructed deep learning network usually have good recognition results,which also makes the current Person Re-identification method based on deep learning basically replace the traditional method.This dissertation mainly takes the multi-granularity network as the starting point,and improves the network by introducing the corresponding modules proposed to extract more discriminative and robust features.The specific research work includes the following three aspects:(1)Considering that the hard partition method of local features used in the current Person Re-identification task,which will cause the problem of misalignment of the obtained local features.Aligned Partition of local feature method for person re-identification is proposed in this dissertation.The center position of the person is learned through the network,and the image is partitioned along the Y-axis to both sides of the image based on this center position,and the local features of the person can be effectively aligned.In addition,for the problem of the difference of triplet loss optimization of local features,the local connection feature is proposed,and optimizing it can make the measurement results of all local features consistent,which helps to gather the final features of person with the same identity in the identification stage.(2)From the perspective of features extraction from deep network,shallow features tend to contain more detailed information,while the high-level semantics of deep features are more abstract.In view of the fact that most of the current research based on deep learning methods ignore the effect of shallow features.Therefore,Feature Fusion Module(FFM)is proposed in this dissertation,which can effectively fuse the features of two different layers.Then,the MultiLayer Features Fusion Network for Person Re-identification(MLFFN)is proposed based on the module,and the final feature obtained contains the information of features of different layers.(3)Given the fact that the attention mechanisms have achieved good results in the research of deep learning tasks,in order to improve the performance of Person Re-identification,this dissertation proposes a Multi-Type Features Network for Person Re-identification(MTFN)by combining various attention mechanisms and extracting features of different granularity.In this dissertation,an improved Convolutional Block Attention Module(CBAM-Pro)is proposed,which is then combined with self-attention module to form a Joint Attention Module(JAM)to extract the global features of person in different areas of attention.Finally,global features of different areas of attention and local features of different granularity are connected to identify person together.Above research works have carried out relevant experimental analysis and comparative verification on three commonly used datasets CUHK03,Market-1501,Duke MTMC-re ID for person re-identification task,and the results prove the effectiveness of the methods in this dissertation.