Recognition Effects of Deep Convolutional Neural Network on Smudged Handwritten Digits

Deep convolutional neural network (CNN) is known to be the first truly successful deep learning approach for image processing and understanding, e.g., the handwritten digits discrimination. However, in real applications such as handwritten zip code recognition, the collected images are commonly with smudged background. In this paper, we study the recognition effects of CNN on smudged digits and compared the results with three-layered perceptron. Experimental results based on MNIST dataset with smudged background (simulated by salt-and-pepper and gaussian noises) show that a drastic decline of recognition accuracy is observed for CNN, suggesting that the extracted features by convolutional operation and max pooling is very sensitive to the noise.