Semantic Guided Deep Unsupervised Image Segmentation

Image segmentation is an important step in many image processing tasks. Inspired by the success of deep learning techniques in image processing tasks, a number of deep supervised image segmentation algorithms have been proposed. However, availability of sufficient labeled training data is not plausible in many application domains. Some application domains are even constrained by the shortage of unlabeled data. Considering such scenarios, we propose a semantic guided unsupervised Convolutional Neural Network (CNN) based approach for image segmentation that does not need any labeled training data and can work on single image input. It uses a pre-trained network to extract mid-level deep features that capture the semantics of the input image. Extracted deep features are further fed to trainable convolutional layers. Segmentation labels are obtained using argmax classification of the final layer and further spatial refinement. Obtained segmentation labels and the weights of the trainable convolutional layers are jointly optimized in iterations in a mechanism that the deep network learns to assign spatially neighboring pixels and pixels of similar feature to the same label. After training, the input image is processed through the same network to obtain the labels that are further refined by a segment score based refinement mechanism. Experimental results show that our method obtains satisfactory results inspite of being unsupervised.