Multi-input topology of deep belief networks for image segmentation

We propose a novel approach for image segmentation by taking the advantages of a 5-layer Deep Belief Network (DBN). DBN composed of multiple layers of latent variables (“hidden units”) which used to extract abstract and robust features for image segmentation. However, it processes images with intricate background, hardly. In order to overcome this limitation, we aim to segment only a few pixels by obtaining their local and regional features at each step. The proposed DBN topology utilizes two different components in two different layers as input to extract local and regional features. Both input components are defined on two different scales of the image. Experimental results show 81.93% accuracy on test images which is the result of providing more information for DBN architecture to learn images with intricate background.