Fast Convergence Detection Algorithm of Image Small Object Based on Distance Intersection over Union

Abstract

Due to low resolution or few features, small object detection has become a difficult problem in the field of image recognition. This paper proposes a fast convergence detection algorithm for small image objects based on distance intersection over union. First of all, EnlightenGAN is used to enhance the image, reduce image noise, and highlight the detection object features. Then, a loss function design of YOLOv5 network based on distance intersection over union is proposed. This method speeds up the gradient regression of the network, greatly shortens the training time of the YOLOv5 network, and improves the detection accuracy. The experimental results using the WiderPerson dataset and the VOC07++12 dataset show that, compared with the traditional YOLOv5 network image detection results, the method proposed in this paper improves AP0.5 by 4.4% and 3.3%, and APs by 6.8% and 2.5%, respectively, which verifies the effectiveness of this method.