In order to solve the problems of low image segmentation accuracy and poor segmentation results for small targets caused by the uneven density of Takifugu rubripes， an instance segmentation method based on improved light SOLOv2 is proposed. Firstly， the structure of deformable convolutional networks （DCN） is optimized by adjusting the receptive field of the convolution using offset parameters.This adjustment enables the receptive field to be closer to the actual shape of the object， leading to better segmentation accuracy.Next，the parameter-free attention mechanism SimAM is fused in the last layer of the residual module to capture more local information in the image， obtain target features at different scales， and optimize the performance of the model for small target segmentation. The experimental results show that the average segmentation accuracy of the improved lightweight SOLOv2 model was improved by 3.7 percentage， and the segmentation accuracy of small targets was improved by 1.4 percentage compared with the original model. After adding both DCN and SimAM attention modules， the segmentation accuracy of the model increased to 65.2%. The results show that the improved SOLOv2 model can improve the detail perception at the boundary， strengthen the model’s ability to extract the features of small target fish stocks， and can be used for accurate instance segmentation in high-density scenarios to achive accurate pixel-level segmentation of Takifugu rubripes.