Further Non-local and Multiple Granularity Network for Vehicle Re-identification

An algorithm that can effectively distinguish different vehicles with high similarity by identifying vehicle photos collected from multiple angles is called vehicle recognition algorithm. This algorithm has been extensively applicable to the region of intelligent transportation and urban computing, but it is always challenging to implement the algorithm. In this paper, an modified feature extraction method is provided, which enhances the traditional nonlocal neural networks and improves its ability to capture the relationship between different positions in the image. At the same time, we adopt a multi-branch global and local information learning strategy, which not merely captures the global features, moreover, those local parts of the feature can be more focused on finer discrimi-nating information in each part of the partition and filtering information on other partitions as the number of partitions increases. Finally, a hybrid relationship between channel feature fusion and channel level is introduced based on this learning strategy. The experimental results show that the MAP and Rank-1 indexes on the VERI-776 mainstream public dataset are 78.9% and 93.86%, which are the highest running scores in this dataset at the present stage, proving that the proposed algorithm is excelled other main stream design.