Underwater Occlusion Target Detection Algorithm Based on Adversarial Attention Mechanism

Abstract

The complexity of underwater environments and the severe lack of occluded target information, make the extraction of sufficient information difficult, resulting in a high omission factor for underwater occlusion targets. To solve this problem, the present study proposes an occluded underwater target detection algorithm based on an improved adversarial attention mechanism. Using Faster R-CNN as an adversary network, the Adversarial Occlusion sample Generation Network (AOGN), which has a competitive relationship with the Faster R-CNN is designed to improve the detection accuracy for occlusion targets. Through a three-stage learning process, AOGN learns how to generate samples that are difficult for the detection network to classify correctly, thereby improving the detection accuracy of the Faster R-CNN for underwater occlusion targets. Subsequently, the Focal loss function is used to increase the proportion of difficult samples in the loss. Finally, to solve the problem of low resolution of underwater images, SE-ResNet50 is used as the backbone in place of VGG16, thereby enhancing the feature extraction ability. Furthermore, multi-scale feature fusion is adopted based on multi-ROIpooling branches to increase the richness of features. The proposed algorithm achieves mean Average Precision (mAP) values of 73.76% and 86.85% and omission factor values of 2% and 7%, on the URPC and underwater common trash datasets, respectively. These results demonstrate that the algorithm effectively outperforms existing detection methods.