Uneven illumination in underwater environments often leads to blurred object boundaries and severe imaging degradation， posing significant challenges for instance segmentation in aquaculture and ecological monitoring scenarios. These issues are particularly pronounced in images of aquatic organisms with complex textures， such as largemouth bass. To address the problems of weak detail representation， ambiguous object contours， and strong noise interference in underwater imagery， this paper proposes an improved underwater instance segmentation method based on the UniInst framework. The proposed approach introduces a Squeeze-and-Excitation （SE） module to enhance channel-wise responses of multi-scale features by modeling global semantic dependencies， thereby improving the discriminative capability of feature representations. In addition， a Spatial Response Modulation （SRM） module is incorporated to further suppress noisy regions and strengthen the preservation of local structural information through spatial response regulation on top of channel semantic calibration. Furthermore， an edge supervision branch is designed， in which a hybrid edge loss combining Binary Cross-Entropy （BCE） loss and the Dice coefficient is employed to guide the model to more effectively exploit boundary structural information during mask prediction. Experimental results demonstrate that the proposed method consistently outperforms the baseline and other mainstream models on instance segmentation tasks， achieving an average precision （AP） of 84.882% and an AP₇₅ of 97.882%. Notably， the proposed method exhibits superior performance in fine structural recovery and high-IoU threshold segmentation accuracy. These results indicate that the proposed approach can effectively improve instance segmentation performance in complex underwater environments， showing strong robustness and promising application potential.