Exploiting Contextual Information of Color Images through Feature Extraction Techniques for Semantic Segmentation

Abstract

In our findings we try to explore spatial context to obtained good results of semantic segmentation. Spatial context has patch-to-patch and patch-to-background. Patch-to-patch context has semantic relationships on visual patterns of two stuffs of a image. Patch-to-background context had semantic relations in image patch and whole background region. In our research we have explored contextual relations based on CRF. CNNs pair-wise potential captures semantic correlation on nearby patches. Researchers in the past used CNN-Sparse CRF.In our model we used CNN-Dense CRF technique to refine our samples to sharpen the object boundaries. CNN-Dense CRF use pair-wise potentials for local smoothness of images. PairWise potentials are log-linear functions for semantic compatibility in image regions. CRF Pairwise is to develop coarse-level prediction. CRF and Potts-model-based pair-wise potential are jointed to obtained good results for semantic segmentation. Keywords: Deep Learning, FCNN, ANN, Adaboost, CRF, SS-Semantic Segmentation.