HPS-YOLOv7: A High Precision Small Object Detection Algorithm

Now deep learn-based object detection can be deployed on drones for criminalinvestigation or military counter-terrorism. Because the proportion of pixels of pedestrians orvehicles in the aerial picture taken by UAV is very small, the probability of detection of smallobjects in the distance is very low or there are omissions. In this paper, the HPS-YOLOv7 algorithmis proposed to improve the detection accuracy of small objects. We have proposed a modifiedhigh-efficiency layer aggregation network for feature extraction, solved the problem that theconvergence of depth models tends to worsen, and lightly processed models with a Bottleneckstructure. We have proposed C-recursively gated convolution, which fully fuses shallow objectsemantic information and enhances the model capacity. To be more helpful for detect small objects,the detection head of 20×20 was replaced by 160×160 detection head, and shallow feature fusionnetwork (SFN) was connected to make up for the information lost by small objects in the deepconvolutional network. Mosaic data enhancement and a priori anchor adaptive adjustment strategyare used in model training to improve the detection efficiency and accuracy. Experimentalevaluation was carried out on VisDrone2019 and Tinyperson data sets respectively. The results showthat mAP increases by 3.0% and 13.29% compared with yolov7 on the basis of IoU=0.5. mAP of0.5≤IoU≤0.95 increased by 1.8% and 3.97%; It is shown that the advantages of HPS-YOLOv7 insmall object detection have certain theoretical value and practical significance.