SEMANTIC SEGMENTATION OF IMAGES USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

In this report, a di erent approach for the semantic segmentation of images is presented. It mainly focuses on the semantic analysis of images that lessen the gap between semantics and low level visual features of images. Here Convolution Neural Network(CNN) with deconvolution layers and Conditional Random Fields (CRFs) is used to get semantic meaning of images. Convolution neural networks are very powerful visual models which produces the hierarchy of features. Fundamental principle of these networks is to aggregate the information over larger image region and collect ne contextual information as the receptive eld of the image represent. Main motive is to design a model that takes input image of a xed size and produces output with e cient activation map of the objects appearing in the image. The model is trained on PASCAL VOC 2007 dataset and getting 70.89% accuracy on training data