EDV-CS-LinkNet: A lightweight semantic segment model of underwater fish school for real-time feeding behaviour quantification in aquaculture

Abstract

Quantifying fish school feeding intensity is crucial for intelligent decision-making in feeding strategies. Real-time and precision semantic segmentation of fish and special distribution characteristics of fish school are essential for feeding behaviours quantification. The loss of spatial details and feature of fish school boundary caused by the uneven illumination and free-swimming fish are the main challenges in available deep convolution network models. In this study, an EDV-CS-LinkNet model is proposed for semantic segment model of underwater fish school to quantify the feeding intensity. It improves the LinkNet method by integrating cross-scale features to make a remarkable balance between accuracy and speed. Specifically, the model employs lightweight encoder-decoder variants (EDV) to extract feature maps and introduces cross-stage skip connections (CS) to encode rich spatial features, addressing under- and over-segmentation issues. Additionally, a special feature fusion module (FFM) is introduced to merge shallow and deep image features. Extensive experimental results demonstrate that the proposed method effectively overcomes the challenges of complex underwater environment and free-swimming fish for underwater fish segmentation. The model achieves an accuracy of 95.3 % IOU with an inference speed of 37 FPS. And, it excels in real-time underwater fish segmentation, enabling precise quantification of feeding intensity in intelligent aquaculture.