CAFM-Unet: Research on Image Semantic Segmentation Based on Transformer

In order to solve the problem that the traditional image semantic segmentation methods have insufficient feature extraction ability when dealing with complex scenes, this paper proposes a CAFM-Unet network for image semantic segmentation. The network adopts the structure of encoder and decoder. In the encoder part, Resnet is used as the backbone network, SegFormer module and CAFM module are introduced, and the feature fusion is carried out. It can complement the modeling ability of the global context and integrate features at different scales and depths. The decoder part adopts the unet structure, introduces the SE attention mechanism, and combines the high-resolution features in the encoder with the low-resolution features in the decoder using the jump connection, which effectively retains the spatial information of the input image and helps the network to recover the image details during the decoding process. The experimental results show that compared with the traditional image segmentation network model, the accuracy of CAFM-Unet network, mpa and MIou reach 94.84%, 93.96% and 88.17%, respectively. It can be seen that CAFM-Unet model can make use of the advantages of convolutional neural network and self-attention mechanism at the same time, solve the limitations of traditional methods in the lack of feature extraction ability, and provide an effective image semantic segmentation method.