Small object detection algorithm based on improved yolov5

Abstract

The detection of small objects is always a difficulty in the field of object detection. This paper proposes an improved yolov5 algorithm to improve the performance of the algorithm for small object detection. In this paper, the BoT3 block is used to instead of the last C3 block in the backbone, which greatly improves the ability of network to feature extraction and fusion ability. Coordinate attention mechanism module is also added to strengthen feature extraction. In addition, EIoU_Loss replaces the CIoU_Loss function to solve the problems that penalizing failure with equal ratio of aspect ratio. At last, in order to better detect small targets in the image, a smaller size prediction head is added. Moreover, before starting the training, the paper uses K-means method to calculate the new anchors. Experimental results show that evaluation indexes of the model have been improved partially. To be specific, the Precision increases 3.8%; the Recall increases 5.9%; the MAP:0.5 increases 6.3%; the MAP0.5:0.95 increases 3.8%.