YOLOX-DC: A Small Target Detection Network up to Underwater Scenes

Abstract

The rapid expansion of underwater target detection technology enables underwater operation robots to better complete underwater tasks. Therefore, a kind of underwater target detection algorithm has been pursued with its high detection accuracy and fast detection speed. One-stage target detection algorithms based on deep learning have achieved marvelous detection speed, but their detection accuracy is poor. In addition to the complex and various underwater scenes, the underwater target is extremely small in the vast ocean background. These reasons lead to the lower detection accuracy of the one-stage detection algorithm in the underwater environment. In this paper, we propose a mechanism to dynamically select feature layer channels, named DC block, so that target features on feature layers with different sizes of receptive field can be better utilized. Moreover, YOLOX, an excellent one-stage target detection network, is called YOLOX-DC after being equipped with the DC block to ensure a faster detection speed and achieve better detection accuracy. On the real-world underwater dataset, for detection and recognition of marine organisms, compared with the advanced two-stage detector Cascade R-CNN, YOLOXDC improves the detection accuracy by 0.9% mAP and the detection time latency is only 3.4% of Cascade R-CNN.