Fully Convolutional Neural Network with Relation Aware Context Information for Image Parsing

Image parsing is among the core tasks in the field of computer vision. The automatic pixel-wise segmentation offers great potential in terms of application adaptability. Traditional convolutional networks have produced better segmentation maps however the research is continued for integration of context information with neural network approaches. In this paper, we propose an image parsing framework that explores the traditional convolutions in fully convolutional networks and learns rich semantic contextual information using the adjacent and spatial modules to generate probability maps. The implicit fusion of the probability maps generated enhances the accuracy of segmentation labels. The proposed framework improves the segmentation accuracy on the CamVid dataset achieving global accuracy of 89.8 %. A comprehensive comparison with state-of-the-art approaches demonstrates that the proposed network exhibits the capability to adapt to the dataset specific information and has the potential to outperform cutting-edge segmentation models.