Context-Based Deep Learning Architecture with Optimal Integration Layer for Image Parsing

Abstract

Deep learning models have been proved to be promising and efficient lately on image parsing tasks. However, deep learning models are not fully capable of incorporating visual and contextual information simultaneously. We propose a new three-layer context-based deep architecture to integrate context explicitly with visual information. The novel idea here is to have a visual layer to learn visual characteristics from binary class-based learners, a contextual layer to learn context, and then an integration layer to learn from both via genetic algorithm-based optimal fusion to produce a final decision. The experimental outcomes when evaluated on benchmark datasets show our approach outperforms existing baseline approaches. Further analysis shows that optimized network weights can improve performance and make stable predictions.