부산대학교 도서관

Aiming at the availability of global features learned by residual network when it was trained to classify images, but was not sensitive to the classification of fine-grained images, a pedestrian re-identification method with enhancing textural features based on parallel residual network was proposed. First, a sub-hard example method was used to deal with the dataset of market1501, by this way the model could converge whiling maintaining good performance. Then structure a parallel network which was used to strengthen textural features, making the network paid more attention to the textural features of images and mixed features from multiple layers for final classification. Experimental results show that sub-hard examples method can improve the accuracy rate of model about 0.2% without changing its parameters, while the value rises to 1.2% if the parallel network is used jointing texture enhancement, with which increasing a few parameters. Improvements were also seen in CUHK03 and MSMT17, two large datasets, at 0.6% and 1.0%, respectively.