YOLO Target Detection Algorithm with Deformable Convolution Kernel

Abstract

In order to solve the problem of inaccurate detection of targets with large scale and shape changes and small targets by the YOLO target detection algorithm, a YOLO target detection algorithm with deformable convolution kernel is proposed. The algorithm adds deformable convolution to the network, and designs three progressive embedding schemes, so that the network can adaptively change the receptive field of feature points according to the shape of the target, thereby extracting features more effectively and improving detection accuracy; by adjusting the backbone network the structure reduces the calculation amount of the algorithm. Tested on the VOC data set, the results show that the algorithm can effectively improve the accuracy of target detection, and the detection speed meets the real-time requirements.