UM-Net: An MLP-based Image Segmentation Network

Abstract

In the field of image segmentation based on deep learning, U-shaped network has always been one of the most popular network frameworks because of its unique structure and wonderful segmentation effect. At the same time, the structure of MLP is simple and compared with CNN and transformer, it introduces less inductive bias and has stronger generalization performance. This paper aims at that the current U-shaped network does not make full use of the rich local information in the deep network structure, and the existing MLP model directly uses fixed weights for feature aggregation and ignores the information of different semantics in different images, focusing only on the structure of the encoder and ignoring the performance improvement brought by the decoder, in order to solve above problems, a new image segmentation network named UM-Net is proposed. It is a network based on convolution and MLP, we add Wave-MLP blocks to the deep structure of the network to make it fully utilize and dynamically aggregate the rich local features in the deep structure, using decoupled dynamic filters to the network upsampling to make it content adaptive. We test our UM-Net on the standard BUSI dataset, the experimental results show that compared with several current mainstream baseline networks, UM-Net achieves better segmentation performance, with F1 score and IoU reaching 80.77% and 68.66% respectively. At the same time, we also conducted different ablation experiments to explore the contribution of each module of the network to the performance.