As the problem of marine debris worsens, underwater garbage detection is of paramount importance for the protection and restoration of marine ecosystems. Due to the complexity of the underwater environment and the blurring of underwater light, the contours of detection targets are unclear, resulting in poor imaging quality and difficulties in distinguishing detection targets. This affects the accuracy and precision of traditional target detection algorithms in underwater garbage detection. After research, this paper proposes a new target detection algorithm called YOLO-PSI. Based on YOLOv8n, a new module called FDP is proposed. The FDP module, through a customized feature focusing module and a feature diffusion mechanism, enables features at each scale to contain detailed contextual information, enhancing feature extraction capabilities. The C2f-PPA module is introduced into the network backbone, mainly for detecting small underwater targets. Through parallel feature extraction, it improves the efficiency of feature extraction and enhances the detection capability of small underwater targets. Focal Modulation replaces the SPPF module in the model to reduce unnecessary computation. The Focal Modulation module increases attention to multi-scale feature maps through weighted fusion, addressing the problem of multi-scale feature map fusion. Experimental results show that the improved YOLOv8-PSI model achieves a mean average precision (mAP@0.5) of 70.3%, which is 3.8% higher than the original YOLOv8n model, making it more suitable for underwater target detection in complex environments.