Improved YOLO network for marine litter detection in a low-carbon environment

Marine litter detection plays this important role in low-carbon environments. Although deep-learning methods have proposed more effective solutions for this task, unclear marine litter photography, severe occlusion of seafloor litter, and poor real-time detection pose great challenges to this task. Based on the above problems, this paper proposes to add Shuffle Attention for local modeling in YOLOv8 to achieve a more complete feature fusion mechanism by effectively reusing multiscale features. Next, the Explicit Visual Center (EVC) module that fuses the multiattention mechanism is used, which can effectively capture the details and contextual information of the target object at different scales. Adaptive adjustment of feature weights and importance. Introducing the Wise-IoU (Weighted Intersection over Union) loss as the bounding box regression loss and using the weight factor and position factor to adjust the position of the prediction box can effectively solve the problem of submarine garbage images being occluded. In this paper, extensive tests are conducted on the Joint Environmental Image and Data Analysis (J-EDI) dataset and a detailed ablation experiment is designed. The results show that the algorithm proposed in this paper achieves improvements in precision, recall and mAP value. Among them, the precision reaches 97.6%, the recall reaches 96%, and the mAP value also reaches 98.5%. Compared with the original network, our improvement has enhanced 2%, 1.1%, and 0.4%, respectively. In addition, the algorithm proposed in this paper achieves a high frame rate of 257 frames per second (FPS), which is an improvement of 40 frames over the unimproved YOLOv8. This means that our improvement is able to process images faster and achieve real-time marine litter detection while maintaining accuracy. Thus, our algorithm is very suitable for real-time application scenarios in low-carbon environments.