Vanilla-Yolo: a lightweight underwater object detector via reparameterization and multi-scale feature fusion

The ocean is rich in mineral, chemical, and marine resources with significant extraction value. Traditional manual underwater operations face challenges such as high risks and low returns. Most previous deep learning-based underwater object detection studies focused on using large networks to improve accuracy, while lightweight underwater detection has received little attention. And many existing algorithms cannot meet the demanding conditions of underwater environments. To address these problems, an optimized lightweight network Vanilla-Yolo is proposed based on VanillaNet and Yolov7. Using VanillaNet as the backbone, feature extraction is enhanced with a series-based activation function. The lightweight multi-feature fusion pyramid with structural reparameterization and multi-scale depthwise separable convolutions is designed to replace traditional convolutions, incorporating adaptive weights to fuse features at different scales and recover global information in underwater images. The improved algorithm demonstrates excellent performance across multiple underwater datasets, achieving 86.6% [email protected] on the DUO dataset. For YOLOV7-tiny, the proposed method reduced the parameter count by 70% while increasing accuracy by 2.1%, achieving a processing speed of 66.4 frames per second (FPS). These results indicate that Vanilla-Yolo can efficiently and accurately detect underwater marine targets, providing an effective method that balances precision and speed.