Lightweight Semantic Segmentation Network based on Attention Feature Fusion

With the popularization of intelligent mobile devices, the lightweight semantic segmentation networks for terminal-oriented have made great progress. However, the low and middle-level features are not be fully used in the networks, and It's often overlooked that low-level noises on high-level features are superimposed, which can cause the important information to be blurred. In this paper, a lightweight semantic segmentation of Attention Feature Fusion Network (AFFNet) is proposed. In this network, a multi-branch structure is adopted to utilize feature information of different stages. An attention feature fusion module is designed, in which the feature information in each stage is weighted with the channel attention and spatial attention to avoid noise superimposition and information coverage in the fusion stage. The loss function of Object Weighted Focus Loss and Cross Entropy Loss (OWFL+CEL) is introduced in the training process to suppress the class imbalance problem. The results show that, with the operation of a single RTX2080 GPU, the mean intersection over union (mIoU) of the method on the Cityscapes dataset is 70.8%, and it reaches 35.7 frames per second(FPS) under high resolution input, which proves that this method has better performance and practical value than similar networks in the tasks of real-time semantic segmentation.