Design And Implementation Of Fine-grained Vehicle Classification System Based On Deep Attention Neural Network

Fine-grained vehicle classification is an important research topic in Intelligent Transportation System.Exact identification for vehicle type information is of great significance to urban traffic management.However,fine-grained vehicle classification still faces many challenges.On the one hand,due to the influence of view,illumination,background and other factors,the same type of vehicles may have great intra-class differences.On the other hand,the differences between different types of vehicles could be very small because vehicles always share similar geometry structure.Deep convolution neural network has achieved good performance in general image recognition tasks,but traditional convolutional neural network can not meet the requirements of fine-grained vehicle classification due to the limitations of feature representation ability.In this thesis,we design and implement a fine-grained vehicle classification system based on the deep attention convolutional neural network.Firstly,in order to eliminate the effect of image background,we improve general object detection methods according to vehicle structure prior knowledge.Our proposed method can detect target vehicles in complex background.Then,we propose a multi-level convolution feature fusion method based on attention mechanism.Our method uses attention module to locate discriminative regions in vehicle images,extracts multi-layer convolution features,and then fuses features in form of feature pyramid.The fused vehicle features can not only focus on the differences of local parts of vehicles,but also combine low-level texture features and high-level semantic features,which has better robustness than before.Finally,we define a loss function based on the hierarchical structure of vehicle labels to further constrain feature vector,which makes the distribution of the feature vector become sparser in feature space.We further implement a fine-grained vehicle classification system with B/S architecture using an open source deep learning framework Pytorch.Expermental results show our system has good classification performance on vehicle images with various illumination and scales.