Relationship aware context adaptive deep learning for image parsing

Abstract

The formation of deep learning architectures is challenged in several aspects, among the major and fundamental steps to develop an effective image parsing network is feature selection. In addition, the exploration of context information in such frameworks is also of prime importance. In this research, a novel architecture that utilizes distinctive feature selection algorithm, and the context adaptive information is proposed. The feature selection algorithm defines the idea of exploring relationship aware information to minimize the similarity among features and select an affluent and optimum set of feature representations. The efficacy of proposed framework is analyzed using several benchmark datasets including Stanford Background, CamVid and MSRC v2. The proposed framework achieves 93.8%, 91.8% and 96.1% global pixel segmentation accuracy on the benchmark datasets respectively. Furthermore, we present a comprehensive comparative analysis with state-of-the-art techniques in the literature. The analysis reveals meaningful refinements in terms of segmentation accuracy.