Existing instance segmentation networks mainly rely on Convolutional Neural Network (CNN) to extract features, but there are limitations in the segmentation accuracy of targets when facing complex environments. To address this problem, this paper proposes an instance segmentation method based on improved SOLOv2, which verifies the effectiveness of using Transformer networks in the instance segmentation task. Specifically, the ResNet-50 backbone network of the SOLOv2 model is replaced with the Swin-Transformer network. Swin-Transformer has the ability to efficiently model long-distance dependencies and integrate global information, thus improving the segmentation accuracy of the model. In addition, this method adopts the Mesh activation function to replace the traditional ReLU activation function, which enhances the feature expression ability of the model while maintaining the computational speed of the model, and further improves the image segmentation accuracy. By training and testing the model on the COCO dataset, the proposed method improves 8% over the traditional method. The results show that the Swin-Transformer network significantly improves the segmentation performance of the SOLOv2 model and provides a new idea for the improvement of the accuracy of subsequent example segmentation tasks.