Tiny Object Detection based on YOLOv5

In view of the poor accuracy of mainstream object detection algorithms in detecting tiny objects, A tiny object detection algorithm based on improved YOLOv5 is proposed. The main feature extraction network of YOLOv5 was modified to generate four feature images to enhance feature extraction of the original input images. Modified the YOLOv5 Neck part, combined with FPN and PANet, carried out feature fusion for four feature maps containing different semantic information, generated better features, and improved the performance of tiny object detection. GIoU loss function was introduced to replace the IoU loss function in the original algorithm to improve the positioning accuracy of tiny objects. Swish activation function was used to replace the original ReLU activation function to better retain target features. The Mosaic data enhancement method was used to enrich the object detection background, and the learning rate cosine annealing attenuation training method was used to dynamically update the learning rate parameters, and the improved YoloV5 algorithm was fused. In this paper, a comparison test is conducted between the original YoloV5 algorithm and CityPrersons data set. Experimental results show that the improved YoloV5 algorithm can effectively improve the detection accuracy of tiny objects.