Person Re-identification Based On Deep Learning

Person re-identification(Person Re-ID)aims to identify and match the person image under different cameras,and it is an image retrieval technique that uses computer vision technology to determine whether there is a query person in the gallery images.With the widespread applications in public safety,smart device monitoring and other fields,this has prompted many scholars to pay attention to pedestrian re-identification.Due to the difference in time,angle,camera parameters and scene changes,during the shooting of different cameras,person re-recognition faces many problems such as low-resolution images,pedestrian posture changes,background clutter,view,light,occlusion and other variations.Therefore,the key issue of the current person re-identification task is still how to extract more expressive and robust features from the target person image.Feature extraction of images in the early days focuses on the underlying hand-crafted low-level descriptors such as color,shape,part,etc.Although the extraction process of the shallow features is simple,its ability to capture picture details is poor.With the continuous expansion of pedestrian datasets and the development of deep learning algorithms,many breakthroughs have been made in the research of person re-identification.Compared with shallow methods,deep learning algorithms have strong recognition capabilities and automatically learn to explore intricate relationships in multidimensional data.This thesis does the following work from the perspectives of supervised learning and unsupervised domain adaptation for pedestrian re-identification based on deep-learning methods:(1)This thesis proposes a network model with self-regulation feature scheme,called SRFnet,to address the problem of the limited ability of the model for learning person features.The proposed SRFnet utilizes global branches to supervise local branches and deeply mine potential features of the network from the perspective of optimization to obtain more distinguished feature representations of persons.Different from methods that add an attention mechanism,the proposed SRFnet can self-adjust feature learning by only adding loss functions between the local and the global features of the network to improve the feature representation ability,and it does not require any additional convolution parameters.Compared with other methods,SRFnet method has made a significant improvement on benchmark datasets for pedestrian re-identification.(2)This thesis proposes a new attention random variation(ARV)module to address the noisy pseudo-labeling issue generated in the clustering process of unsupervised domain adaptation for person re-identification.Specifically,it is a parameter-free random transformation module that expands the variability and complementary of the network by randomly enhancing the feature mapping unit.The ARV module is integrated into the MMT framework named attention augmented mutual network(AAMN)to produce random differences between two co-networks and avoid network convergence to the same noise.Experimental results show that the m AP of AAMN is higher than the baseline for unsupervised domain adaptation task.