Feature blend-based multiscale semantic segmentation

Semantic segmentation is a key task in computer vision, which splits an image into different semantic blocks by classifying each pixel in the image for subsequent image parsing. Although deep learning has made significant progress in semantic segmentation, existing algorithms still have limitations in capturing global information of large-scale targets and detailed information of small targets. In addition, the commonly used down-sampling operation may lead to feature loss, which in turn affects the accuracy of edge localization. In order to solve these problems, a feature blend based multiscale semantic segmentation model is proposed in this paper. Specifically, we design the FeatureBlendNet module, which enhances the information representation of multi-scale targets in the same category by fusing the original image features and the intermediate features in the backbone network using the relation matrix. And we also designedtheMultiLayerOptimizedDecoder module, which fuses the feature maps of different scales in the decoder, enhances thefeature expression of the network, expands the receptive field, and improves the expression of edge details. Experimental results show that the proposed model has strong adaptability and better segmentation performance in complex scenes. A large number of ablation comparison experiments on the Cityscapes benchmark dataset confirm the superiority of the proposed method.