Abstract

Deep neural networks have recently been achieving state-of-the-art performance on computer vision tasks, especially visual classification problems. But the works mostly use clear and non-blurred images as inputs. In this respect, we propose the neural network adapted to recognize low-resolution digit images. Our network is consisted of proposed convolutional and deconvolutional layers adapted to LeNet, which is convolutional neural network to have 99% performance for recognizing high-resolution digit images. We further show that traditional up-scaling method and state-of-the-art super-resolution method. Our results recognition rate achieve highest performance on each size of inputs. Interesting thing is the feature map induced by our network is unrecognizable to human, but is more recognizable to neural network. This result implies that the difference between human vision and neural network is quite big.

[PDF]