Class-Wise Fully Convolutional Network for Semantic Segmentation of Remote Sensing Images

Semantic segmentation is a fundamental task in remote sensing image interpretation, which aims to assign semantic label for every pixel the given image. Accurate still challenging due complex distributions of various ground objects. With development deep learning, series networks represented by fully convolutional network (FCN) has made remarkable progress on this problem, but accuracy far from expectations. This paper focuses importance class-specific features different land cover objects, and presents novel end-to-end class-wise processing framework segmentation. The proposed FCN (C-FCN) shaped form an encoder-decoder structure with skip-connections, encoder shared produce general all categories decoder process features. To be detailed, transition (CT), up-sampling (CU), supervision (CS), classification (CC) modules are designed achieve transfer, recover resolution feature maps, bridge modified decoder, implement classifications, respectively. Class-wise group convolutions adopted architecture regard control parameter numbers. method tested public ISPRS 2D labeling benchmark datasets. Experimental results show that C-FCN significantly improves performances compared many state-of-the-art FCN-based networks, revealing its potentials accurate images.