Semantic Segmentation of Human Faces Using Convolutional Neural Network

Abstract

In this thesis, we study the human faces semantic segmentation topic using convolutional neural networks. We apply two convolutional neural network structures. These two proposed networks are trained and tested by image dataset that is generated by ourselves. Using simulating data, we are able to know the exact true location of faces in images. We evaluate the first convolutional neural network with different amount of images, and with various of output image sizes. Besides, we change the number of stacked convolutional layers to investigate the effect of deep learning network depth to its accuracy. Results show that the proposed convolutional neural network is suitable and efficient for human faces segmentation task. We also find that increasing the image number and depth of convolutional neural networks, as well as decreasing the output image size, will increase the segmentation accuracy. The second convolutional neural work structure is fixed and more complex than the first one. This network also achieves a higher accuracy than the former network architecture. However, due to its complexity, it requires multiple of time to train and evaluate than the former network.