RESUMO
Limited by GPU memory, high-resolution image segmentation is a highly challenging task. To improve the accuracy of high-resolution segmentation, High-Resolution Refine Net (HRRNet) is proposed. The network is divided into a rough segmentation module and a refinement module. The former improves DeepLabV3+ by adding the shallow features of 1/2 original image size and the corresponding skip connection to obtain better rough segmentation results, the output of which is used as the input of the latter. In the refinement module, first, the global context information of the input image is obtained by a global process. Second, the high-resolution image is divided into patches, and each patch is processed separately to obtain local details in a local process. In both processes, multiple refine units (RU) are cascaded for refinement processing, and two cascaded residual convolutional units (RCU) are added to the different output paths of RU to improve the mIoU and the convergence speed of the network. Finally, according to the context information of the global process, the refined patches are pieced to obtain the refined segmentation result of the whole high-resolution image. In addition, the regional non-maximum suppression is introduced to improve the Sobel edge detection, and the Pascal VOC 2012 dataset is enhanced, which improves the segmentation accuracy and robust performance of the network. Compared with the state-of-the-art semantic segmentation models, the experimental results show that our model achieves the best performance in high-resolution image segmentation.