Find in Library

Aiming at the problem of low visibility of underwater environment, which leads to the leakage of small target detection and low accuracy, this paper proposes an improved algorithm USSTD-YOLOv8n (Underwater small-size target detection YOLOv8n) based on YOLOv8n. First, CARAFE is adopted as anew up-sampling method to achieve more correct feature reconstruction under low underwater visibility. Second, Context Guided Block (CG block) is introduced to replace part of the convolutional structure, which makes USSTD-YOLOv8n have stronger feature extraction capability. Finally, Inner-CIoU is adopted as the loss function to improve the generalization ability of USSTD-YOLOv8n, to obtain more correct detection results. To verify the robustness and accuracy of the model, a new experimental strategy is used to perform one set of ablation experiments and three sets of comparison experiments on the URPC2018 and URPC2020 datasets, the mAP @ 0.5 was 0.7670,0.7910 and 0.7044, compared to the YOLOv8n algorithm, map@0.5 increased 0.0260, 0.008 and 0.007. It is proved through four sets of experiments that USSTD-YOLOv8n has better detection performance in underwater small target detection task.