Improvement of semantic segmentation algorithm based on ResNeSt network

This paper proposes an improved semantic segmentation algorithm based on the split attention network resnest. The original icnet semantic segmentation model has some defects in the segmentation of foreground objects. In view of these problems, we optimized the icnet model with resnest network and CBAM convolution attention module. On the one hand, we replace the backbone network in the original model with resnest network, which enhances the feature extraction capability of the model. On the other hand, we introduce the CBAM convolution attention module to improve the segmentation ability of the model to the target in the foreground region. Without affecting the inference rate of the original model, the training effect of the improved idea proposed in this paper on the cityscapes open data set is improved separately from 71.0% to 71.3% and 71.9%. Respectively compared with the MIoU of the baseline model, the MIoU of the icnet model combined with the resnest backbone network and the CBAM convolution attention module on the cityscapes open data set is increased from 71.0% to 72.1%, and the segmentation effect of the foreground area target is significantly optimized in the visualization images, which has certain positive significance for the application of the semantic segmentation algorithm in the field of unmanned driving.