一种在低质量图像上提高字符识别率的深度学习框架

Improves Character Recognition Rate on Low Quality Images by a Deep Learning Framework

论文为了解决低质量图像给识别任务带来的困难,构造了一个由图像增强网络(EnCNN)和手写体数字识别网络(LeNet-5)组成的低质量图片识别框架。将图像增强网络嫁接在识别网络前,并使用论文提出的策略进行模型学习。使得低质量图像在被识别前图像质量得到较大的改善,最终实现低质量手写体图像识别率的提高。实验部分将论文提出的方法和在单纯使用低质量图像或高清图作为训练集进行训练的方法进行了对比,实验表明在低质量图像上,论文提出的方法有更高的数字识别率,且有更强的泛化能力。

Based on the convolutional neural network model for handwritten numeral identification,this paper constructs an image framework with image enhancement network and a handwritten digital identification network in order to solve the difficulties caused by low quality image. The image enhancement network is grafted to identify the network,and the model proposed by this paper is used to model the learning. Which makes the image quality of the low quality image be improved greatly before the recognition task is carried out,and finally the recognition rate of the low quality handwriting image is improved. In this paper,the method proposed in this paper is compared with the method of enhancing the training set. Experiments show that the proposed method is better to other methods and has stronger robustness.