Semantic Segmentation of RGB-NIR Images with Error-Correcting Output Codes

Scene understanding is strictly linked to image semantic segmentation, which is the process of associating each pixel of an image with a label, such as sky, clouds, road, building. This paper proposes a new semantic segmentation framework, in which Error-Correcting Output Codes (ECOC) are used to decompose the multiway classification problem into multiple binary classification subtasks. The binary output results are then converted into final class labels following a decoding table established at the beginning of the classification procedure. As part of the recognition framework, color descriptors and high-level visual features are extracted to represent the appearance of the patch surrounding each pixel of interest. The proposed method is validated on an image database containing RGB and Near-Infrared (NIR) imaaes.