GCP-YOLO: a lightweight underwater object detection model based on YOLOv7

With the rapid development of technology, underwater biological detection tasks are generally conducted using mobile devices. This paper proposes the GCP-YOLO model based on YOLOv7 to address the challenges of deploying large detection models on mobile devices in the field of underwater object detection, particularly the issues of difficulty in underwater object detection and resource constraints. First, the GhostNetV2 module is used to make the ELAN module lightweight in the Neck part, reducing the model's parameter count and computational complexity. Second, to address potential issues such as feature loss and low accuracy when collecting feature information in the lightweight network, we incorporate the CA Attention module after the improved ELAN module to prevent feature loss caused by the lightweighting process. Finally, we perform pruning on the overall improved model with a pruning rate of 50%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document}, further reducing the model's parameter count and computational complexity. Compared to the YOLOv7 model, the GCP-YOLO underwater object detection model reduces the parameter count and computational complexity by a factor of 4, while improving accuracy by 2.8%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document}.