Scene Labeling Based On Convolutional Neural Networks

This paper considers the application of convolutional neural networks(CNNs) in the face-background classification recognition, the hand written digits recognition, the images classification recognition and scene labeling. By combining the CNNs with different feature extraction filter layers and other optimization methods, we get the different image information with multi-scale convolution kernels in the same image. And a multi-scale CNN model which solve scene labeling problem as we exploring improves the accuracy of scene labeling by applying the method to Stanford Background dataset and SIFT Flow dataset. The main work is as follows:(1) By using the method of CNN with different feature extraction filter layers, we conduct the face-background classification, the handwritten digits recognition and the images classification. At the same time, we compare the results of three experiments with other model and study the structure of CNN. In face-background classification, the average classification accuracy reached 99.785%; In handwritten digits recognition, the highest accuracy is reached by the model CNN+SGD which is 96.7%, which is 8.1% higher than the model MLP+SGD; In multi-class images classification, the highest accuracy is reached by the model CNN+SGD which is 66.7%, which is 52.3% higher than the model MLP+SGD. At the same time, the experiment shows the way of ReLU and pooling of CNN to extract the image features.(2) As for solving the scene labeling, we propose a multi-scale CNN which use the multi-scale convolution kernel to extract the image feature to obtain the information of same images in different scales and put it into scene labeling. In the scene labeling of Stanford Background, we do some processing which put multi-scale of test image into CNN and get 33.5% higher accuracy than without. The multi-scale CNN is applied to scene labeling with SIFT Flow which improves the accuracy by 36.3%. The method verifies that increasing the scales of test images can improve the accuracy in solving the scene labeling. The experiments on scene labeling also indicate that ReLU and Dropout can improve the generalization ability of CNN.