Multi-Layer Feature Fusion for Small Target Detection Based on Yolov5

Due to the low visibility of small targets, small target detection (STD) is considered one of the most challenging tasks in the object detection framework. The term “small target” refers to objects that occupy a small proportion of the input image. For example, the widely used MS COCO dataset defines objects with a border size of 32×32 pixels or smaller in a typical 480×640 image as small targets. In contrast, other datasets have their definitions, such as objects that make up 10 % of an image. Small targets are often missed or detected by misplaced borders and sometimes incorrectly marked. The main reason for the inaccuracy of STD localization is the degradation of subsequent space due to the limited information provided by input images or video frames superimposed on the deep network layer. Since small targets often appear in various application fields, such as pedestrian detection, medical image analysis, face recognition, traffic sign detection, traffic light detection, and ship detection, it is worth studying how to improve the performance of modern deep-learning STD techniques. Therefore, we proposed an improved Yolov5-based small targets detection method in this paper. Firstly, based on the original anchor box, a group of smaller anchor boxes is added to detect small targets in the shallow feature map. In addition, we split the image into different small areas for detection and then converted them into standard image sizes. This method is straightforward and effective to help us focus on small targets in an image. Finally, we upsample the deep feature map and concatenate it with the shallow feature map to obtain a larger feature map for STD. as shown in Fig. 1. Experiments show that our model achieves a mAP of 45.1 % on the VisDrone dataset, which far exceeds the 35.7 % detection score of the Yolov5 model.