A two-column convolutional neural network for facial point detection

This paper proposes a two-column convolutional neural network algorithm for face point detection, which proves better performance than single-column. Because of the deep level structure, global texture features are extracted at higher layers of the network, and the results of keypoint location remains high accuracy. In the first column, we take the R, G, B channel component of original image as input to train the neural network respectively and calculate the average location results. In the second one, we use Sobel operator to extract first-order derivative feature of original image and regard it as the input to train another convolutional network. The location results of two columns above are fused with appropriate proportion to ensure better results. Experimental result on LEW dataset shows the performance of the proposed parallel structure better than single convolutional network, and it stays robust due to occlusions, large pose variations and extreme lighting.