Lightweight LIS-YOLO Model for Multi-target Fish Instance Segmentation

Abstract

Accurate segmentation of underwater objects laid the foundation for studying aspects such as the behaviour and biomass of aquatic animals. However, existing underwater instance segmentation algorithms often lacked sufficient robustness when facing underwater environment-specific interferences, such as suspended particles, colour attenuation, and background noise, and enhancing the generalization ability of lightweight models in unstructured and dynamic underwater scenes remained a key challenge. To this end, a lightweight instance segmentation lightweight instance segmentation YOLO v8 (LIS-YOLO) model was proposed, which can effectively segment sturgeon, juvenile bass, and adult bass from different shooting angles, and a real-time underwater object segmentation system was developed based on PyQt5. Firstly, a lightweight and high-precision C2f-Faster-EMA module was designed, in which the original complex C2f module was replaced with a more lightweight C2f-Faster module, and an efficient multi-scale attention mechanism was integrated to improve the feature extraction capability for small-object fish. Secondly, Wise-IoU was introduced into the improved model to reduce harmful gradients caused by low-quality samples, thereby enhancing the model’s segmentation capability in complex environments. Finally, a real-time multi-object instance segmentation system for underwater objects was developed by using a graphical user interface and the PyQt5 framework, enabling the visualization of different fish species. The experimental results showed that the LIS-YOLO model achieved precision, mean average precision, floating-point operations, and frame rate of 97.2%, 95.9%, 3.60×1010, and 127f/s, respectively. The number of model parameters was compressed to 9.0×106, accounting for 76.3% of the original model. This research result not only provided an accurate and lightweight instance segmentation model for underwater object recognition but also explored the effectiveness of fish segmentation from different shooting angles, offering practical application value for improving the level of intelligent aquaculture.

Keywords: deep learning, instance segmentation, multi-object fish, YOLO v8, PyQt5, lightweight model