Abstract: This paper proposes an image semantic segmentation method based on feature fusion and self attention mechanism, and designs feature fusion module, self attention module, enhancement module, global spatial information fusion module and loss function. The feature fusion module fuses all components from multiple images and executes them through the self attention mechanism. Self attention module can effectively capture remote context information. The enhancement module aims to enhance the input image to obtain more diversified features. The global spatial information attention module has only linear complexity relative to the image size, which can bring significant improvement effect. The loss function is used to optimize the model, and the classification result of each pixel is optimized to the nearest real value. The experimental results show that the proposed method can significantly improve the performance of PASCAL VOC 2012 dataset, COCO Stuff 10K dataset and ISIC 2018 dataset, and has been verified on three datasets. The experiment also verifies the advantages of the method in this paper by comparing self attention, reasoning speed and ablation experiments.