Image segmentation by a new type of convolutional neural network

Abstract

In computer vision, image segmentation is a process that partitions a digital image with objects into segments of the objects and background. The goal of image segmentation is usually to separate the images of the objects for identification. The segmentation process assigns the same color to all the pixels in the image of an object. Notable applications include street scene understanding for self-driving cars and medical imaging for radiograph analysis. The convolutional neural network (CNN) for image segmentation is known to have inadequate accuracy as compared to the accuracy of the radiologists. A new type of CNN, called Segmentation CNN, is proposed in this theses. Preliminary numerical experiments on the well-known benchmark dataset, MNIST dataset, show the segmentation CNN outperforms the standard CNN by a large margin consistently. More work is called for to develop the Segmentation CNN.