RMNET: A Residual and Multi-scale Feature Fusion Network For High-resolution Image Semantic Segmentation

Abstract

High-resolution remote sensing images have high clarity and provide significant support for urban planning, resource management, environmental monitoring, and disaster warning. Semantic segmentation accurately helps extract the boundaries of objects, thereby increasing the application value of scene understanding. Traditional encoder-decoder architecture networks lack multi-scale information fusion and fail to capture precise multi-scale semantic information, when segmenting targets at different scales. Additionally, these semantic segmentation networks have inadequate handling of class-imbalanced data, resulting in unsatisfactory classification results and final segmentation effect. This paper proposes a semantic segmentation network based on residual blocks and multi-scale feature fusion. Building upon the U-Net network, we design residual modules and multi-scale feature fusion modules to extract information-rich feature maps. Then, the multi-scale feature fusion module is used to interpolate and upsample the obtained feature maps, which are then concatenated with feature maps at the same layer, resulting in a novel fusion feature map. In experiments, the performance of the proposed model surpasses U-Net with improvements reaching 6.06% for MIoU. The introduced network identifies complex land features including dense distribution of objects, small objects, large differences in object characteristics and complex background effectively preserves and restores feature information by incorporating the multi-scale feature fusion module, achieving higher precision segmentation results and providing rich multi-scale and spatial information