Traffic violation image processing method based on improved YOLO-v4

Aiming at the situation that vehicles in traffic violation pictures are forced to violate regulations or drive under the command of traffic police, we propose a method for further information extraction of traffic violation images based on the improved Y0L0-v4 network. According to the Y0L0-v4 detection results, make basic judgments on the intersection situation and vehicle status in the traffic violation pictures, and provide further relevant information for subsequent manual screening to reduce the burden of manual work. In view of the characteristics of the target to be detected in the traffic violation image, the Mosaic data augmentation in Y0L0-v4 is replaced by Mixup data augmentation. Due to the low detection accuracy of small targets such as traffic lights and traffic piles in traffic violation images, a variety of attention mechanisms are compared, and finally, the Coordinate Attention attention mechanism which has a better effect on improving the traffic violation image datasets and the fewer parameters is selected. The Depthwise Over-parameterized convolutional layer(DO-Conv) is used to replace the basic convolutional layer to extract input image information more efficiently and further improve the accuracy of network detection. Finally, We propose a lightweight Y0L0-v4, which greatly reduces network parameters and improves the speed of network detection while meeting the detection requirements.