CNN-Based Cascade with Skipping Connections for Semantic Segmentation

Abstract

Convolutional neural networks (CNNs) have produced significant improvements in semantic segmentation. They can collect relevant contextual information directly from the RGB images via stacks of convolutional and subsampling layers. In this paper, semantic segmentation is solved using a novel cascade of CNNs. The proposed method employs skipping connections for combining the input images with intermediary results, thus enabling successive corrections of the label maps. The cascade can easily integrate additional corrective algorithms, as exemplified for a graph-cut algorithm with confidence-dependent weight cues. The design allows a separate training of each component network, with reduced computational resources. The performance of the proposed approach is investigated for RGB images acquired by a wearable assistive device, in the framework of an application assisting the navigation of visually impaired persons. The experimental results indicate that the component CNNs can gradually improve the accuracy of the semantic segmentation.