Dense image labeling using deep learning
Amirul Islam · Mspace (University of Manitoba) · 2017
Recently there has been remarkable success in pushing the state of the art in dense image labeling tasks. Most of the improvements are driven by employing end-to-end deeper feed-forward networks. First, we propose a dense image labeling approach based on Deep Convolutional Neural Networks coupled with a support vector classifier. However, in many cases precisely detecting smaller and thinner object details require representation of fine details. To overcome this limitation, we propose end-to-end encoder-decoder networks that initially make a coarse-grained prediction which is progressively refined to recover spatial details. This is achieved by gate units proposed in this thesis, that control information passed forward in order to resolve ambiguity. Furthermore, we propose an end-to-end salient object detection network that employs recurrent refinement to generate a saliency map in a coarse-to-fine fashion. Experimental results demonstrate the superiority and effectiveness of our proposed approaches.