Research On Vehicle Detection And Attribute Recognition Methods Based On Deep Learning

The frequent appearance of fake license plates and the complexity of real traffic scene limit the role of license plates in vehicle identification.As a part of vehicle information,vehicle attributes become a significant criterion for vehicle identification During various vehicle attributes,the features of vehicle brands and vehicle types are fine-grained and difficult to be extracted.As the previous work of vehicle attribute recognition,vehicle detection needs to provide complete vehicle images for vehicle attribute recognition,but incomplete images are likely to cause the reduction of recognition accuracy.Compared with the traditional methods of vehicle detection and vehicle attribute recognition,the methods based on deep learning perform better However,there still are some problems like incomplete detection and poor extraction of fine-grained features.To solve the problems above,the thesis studies the methods of vehicle detection and vehicle attribute recognition separately(1)A method of vehicle detection based on self-adopted anchor box is proposed in the thesis.The two-layer fully connected network is used to perform regression on the prior values and the anchor boxes with hyperparameters are generated.Moreover,another regression on the coordinate of anchor boxes is operated to create bounding boxes by YOLOv3.Our model makes the coordinate prediction more accurate than the other ones after two regressions.Meanwhile,it provides complete vehicle images for vehicle attribute recognition(2)A method of vehicle attribute recognition based on attention mechanism is proposed in the thesis.B-CNN model highlights the local area weights to enhance the feature map by cross production.Based on attention mechanism,an APN structure is added to find the attention area.The attention area is enlarged for the fine-grained classification of vehicle brands and types.After feature extraction,a multi-lab el classifier is designed to classify different kinds of vehicle attributes.The experiment reveals that our method performs better than the other ones on all experimental datasets.