Research of Person Re-identification Based on Deep Learning

Person re-identification is widely considered as a sub problem of image retrieval. The main method is giving a query image and retrieve the same identity pedestrian image from the corresponding image library. In this process, there is often a mismatch phenomenon, which leads to the decline of the first matching rate of Person re-identification. In order to alleviate this problem, this paper has made progress in the feature extraction training set of deep convolution neural network One step research, improve the algorithm, improve the matching degree of rank-1 and map. We analyzes the development of deep learning related methods in Person re-identification in recent years firstly, summing up and integrating some excellent algorithms and related network models in this field, and then solves the problem based on this, using feature map for horizontal segmentation, and then calculates the loss method to enhance the learning rate, which improves on the original baseline. Design a strong baseline, add a layer of linear and batchnorm layer, then connect the classifier, remove the last two FC's bias. Then we try to use a global feature which combines the simple training skills to test. When training the input image, we use rectangle to intercept, which can improve the matching degree of the image. Finally, the development direction is prospected and the future development focus is discussed.