Dual-branch underwater object segmentation method based on image enhancement

Light absorption and scattering in water easily cause image color distortion and contrast degradation, reducing segmentation accuracy and anti-interference ability for underwater targets. To address this issue, an underwater target semantic segmentation framework was proposesed. Firstly, an image enhancement module was applied to restore optical degradation, correct color deviation, and provide high-quality feature inputs. Secondly, a four-stage dual-branch network integrating convolutional neural network (CNN) and transformer was constructed to extract local details and global semantic features, and a spatial fusion module was adopted to achieve multi-scale feature complementarity. Finally, a feature aggregation module was introduced to optimize contour features and achieve high-precision segmentation. Comparative experiments were conducted on four underwater datasets against nine mainstream segmentation models. On RUWI dataset, the proposed method achieved an average class accuracy of 94.35% and a mean intersection over union of 88.42%. Experimental results demonstrate that the proposed method significantly outperforms existing mainstream algorithms in both segmentation accuracy and model robustness, can stably chieve accurate target segmentation in complex underwater scenes.