Lightweight Underwater Target Detection Algorithm Based on Dynamic Sampling Transformer and Knowledge-Distillation Optimization

Abstract

Underwater robot perception is a critical task. Due to the complex underwater environment and low quality of optical images, it is difficult to obtain accurate and stable target position information using traditional methods, making it unable to meet practical use requirements. The relatively low computing power of underwater robots prevents them from supporting real-time detection with complex model algorithms for deep learning. To resolve the above problems, a lightweight underwater target detection and recognition algorithm based on knowledge distillation optimization is proposed based on the YOLOv5-lite model. Firstly, a dynamic sampling Transformer module is proposed. After the feature matrix is sparsely sampled, the query matrix is dynamically shifted to achieve the purpose of targeted attention modeling. Additionally, the shared kernel parameter convolution is used to optimize the matrix encoding and simplify the forward-propagation memory overhead. Then, a distillation method with decoupled localization and recognition is designed in the model-training process. The ability to transfer the effective localization knowledge of the positive sample boxes is enhanced, which ensures that the model maintains the same number of parameters to improve the detection accuracy. Validated by real offshore underwater image data, the experimental results show that our method provides an improvement of 6.6% and 5.0% over both baseline networks with different complexity models under the statistical index of detection accuracy mAP, which also suggests 58.8% better efficiency than models such as the standard YOLOv5. Through a comparison with other mainstream single-stage networks, the effectiveness and sophistication of the proposed algorithm are validated.