基于卷积神经网络的自然场景门牌多数字识别研究

Multi-Digit Number Recognition from SVHN in Natural Scene Based on Convolutional Neural Network

自然场景门牌号码由于图像模糊、光照不均匀、弱光照等情况下,产生严重的畸变,导致字符识别难以取得理想的效果,识别任意长度的字符更是一个难题。用卷积网络提取特征,构建一个识别单数字的卷积神经网络;在不分割字符的情况下,用循环网络生成字符序列,构建一个识别多数字的深度卷积神经网络。使用卷积网络和循环网络融合的网络结构,在SVHN数据集上进行验证,精度方面取得较好的效果,尤其是单数字门牌号码的识别率95.72%,多数字门牌号码的识别率89.14%。

Due to blurred image, uneven illumination, weak illumination, etc., the natural scene number is difficult to achieve desired result. It is more difficult to recognize string of arbitrary length. Proposes the convolutional network and the recurring network, the former is to extract the features, and the latter is to generate the final character sequence, constructs a convolutional neural network that recognizes single digit and a deep convolutional neural network that identifies multi numbers without splitting the characters. By validating it on the public SVHN dataset, the recognition rate of single number is 95.72%, the recognition rate of multi-digit number is 89.14%, and the accuracy is better than other methods.