Augmented-High Resolution Network for Image Segmentation

Convolutional Neural Network (CNN) based architectures have shown superior performance in computer vision. But in semantic segmentation, CNN based models suffer from loss of contextual information due to reduction in resolution. High Resolution Network (HRNet) addresses this issue by preserving parallel processing layers at high resolution, but it also suffers from loss of information due to low and high resolution fusion. In this work, a unique approach to do semantic segmentation of images using Augmented HRNet (Au-HRNet) is presented. To prevent information loss during fusion of low resolution with high we have added channel and spatial attention blocks. These blocks are added at each stage and within the fusion module. The augmented HRNet output is then passed to contextual representation segmentation header. By enriching pixel through weighted aggregation using object contextual representation, pixel’s label matching to the object category is ensured. Through this approach, we have achieved improved results on widely used datasets like PASCAL VOC+AUG and PASCAL CONTEXT.