Credential Image Text Recognition Based On Deep Neural Network

With the widespread application of current smart identity authentication,more and more users upload credential images taken by smart phones.This puts great working pressure on Internet companies that need to be authenticated and the departments of government that need to recognize and record the credential image text.Automatic and accurate identification of credential image text information,to replace the cumbersome and error-prone manual processes,shorten the workflow and improve the efficiency of the work has a positive meaning.The traditional Optical Character Recognition(OCR)technology is mainly for high definition scanning images,this method requires that the recognized image has a clean background,uses a standard printed matter and has a higher resolution.However,there are many problems in natural scenes,such as large text background noise,irregular text distribution and natural light influence.Therefore,the recognition rate of OCR technology is not ideal in the actual natural scene.There are also researchers using SVM and deep neural network classifier to complete the recognition of image characters in natural scenes,but there are also problems of low accuracy and high time complexity in real natural scenes.Therefore,in view of the above problems,this paper applies depth neural network and other techniques to study the credential image recognition method under natural scenes.The main creative work of this paper is as follows:(1)In view of the current time complexity of image character segmentation algorithm and the improvement of segmentation accuracy,this paper improves the algorithm of image character segmentation based on the traditional Selective Search algorithm using the dyadic wavelet transform.First,the image character projection histogram is used to detect the character by using the dyadic wavelet decomposition,and then use the improved Selective Search algorithm to segment the image characters to generate the input character candidate sets for the subsequent deep neural network image recognition model.Experiments show that the proposed image character segmentation algorithm is better than the IB algorithm and the LP algorithm,which improves the segmentation accuracy and reduces the segmentation time.(2)In view of the gradient vanishing for deep neural networks with deeper layers,which cause the problem of low accuracy and longer training time.In this paper,we use the deep energy model to improve the image recognition method based on the stochastic depth neural network.Firstly,we use the Bernoulli distribution probability to choose the current neural network layer,to prevent the phenomenon of gradient dispersion in the training process.Then,in the process of back propagation,energy function is designed to optimize the loss function of neural network.Experiments show that the proposed method based on stochastic depth neural network improves the recognition accuracy of the credential image text and reduces the running time.