Segmentation of underwater fish in complex aquaculture environments using enhanced Soft Attention Mechanism

Underwater fish segmentation technology serves as a crucial foundation for extracting aquatic biological information. However, due to intricate and fluctuating underwater environments, existing segmentation models fail to precisely focus on key image regions. Based on this, the paper developed an underwater fish segmentation model, Receptive Field Expansion Model(RFEM), by enhancing soft attention performance (More attention is directed to fish regions when processing fish pixels). This paper tests ten different attention mechanisms and selects the attention mechanism with better performance indicators to improve it and form an RFEM model. This paper uses two underwater fish data sets to verify the proposed model. The experimental results show the segmentation mean intersection-over-union ratio (MIoU) of RFEM based on dilation convolution reached 88.37%, and the mCPA reached 93.83%, Accuracy reached 96.08%, and F1-score reached 93.74%. It can provide solid technical support for intelligent monitoring such as body length measurement, weight estimation of underwater fish.