An enhanced underwater fish segmentation method in complex scenes using Swin transformer with cross-scale feature fusion

The segmentation of underwater biological images holds significant importance for achieving intelligent aquaculture. However, the complexity and variability of the underwater environment result in images that are often blurred, exhibit color distortion, and have imbalanced contrast, leading to low adaptability of segmentation models and affecting segmentation accuracy. To address these challenges, this paper proposes an underwater biological image segmentation model based on the Swin transformer. Firstly, a multi-branch parallel convolutional residual fusion mechanism is introduced to extract detailed features from the original image, enhancing the model’s capability to capture fine and edge details. Secondly, an enhanced global–local feature fusion module is constructed to strengthen the extraction of both global context and local features, effectively combining local and global dependencies. Lastly, a channel attention mechanism based on a parallel residual structure is proposed, which utilizes multiple receptive fields to emphasize salient features. The features are then fused with multi-scale features through element-wise multiplication, enhancing the transformer’s ability to extract foreground features. Experiments were conducted on the underwater images segmentation dataset (UISD) and a large-scale fish dataset to validate the effectiveness of the proposed underwater biological image segmentation model. The model achieved a mean intersection over union (MIoU) of 92.73%, with improvements of 2.82%, 2.86%, and 2.54% in accuracy (Acc), class pixel accuracy (CPA), and F1-score, respectively, compared to existing classical network models. The experimental results demonstrate that the proposed model exhibits superior performance in underwater biological image segmentation tasks. Additionally, as the number of segmentation categories increases, the model shows good predictive results compared to other classical segmentation models, indicating its advantage in multi-category segmentation tasks. The code and models are publicly available at https://github.com/se-code82/code.git.