Hybrid attention network for semantic segmentation

Abstract

Deep neural networks have made significant improvements in pixel-level semantic segmentation. However, the existing semantic segmentation algorithm still faces the problem of weighing between the accuracy and calculation cost of the segmentation. In response to this issue, this article proposes a hybrid network structure (HAM), looking for a balance point in the calculation accuracy and calculation speed. In this method, we construct a dual-attention module. The role of this module is to guide high -level characteristics through underlying characteristics to obtain more context information. Among them, the shape flow branch retains low-level space details, and semantic flow branches capture senior context information. These two branches are fused to strengthen information dissemination between different levels, thereby achieving higher segmentation accuracy. Experiments on the dataset show that this method achieves higher accuracy and speed under relatively small parameters. Compared with other real-time semantic segmentation methods, our network has achieved good compromise between parameters, speed, and accuracy.