Chinese Invoice Character Recognition Scheme In Complex Background

Receipts play an important role in our daily lives.Converting receipt images into text data that can be processed by computers is an important application scenario of OCR technology.Existing methods usually utilize image preprocessing to obtain grayscale images,which are then input into deep neural networks for recognition.This approach shows excellent performance in most cases.However,in some cases,due to seals and bill templates,the background of the character's backplane may be overly complicated,such that the existing solutions would perform poorly.The main reason is that it is difficult to preprocess the image in these cases,and intractable to have an universal solution for receipts in different formats.To combat these aforementioned challenges,this article proposes a novel method for character recognition on specific layout receipts,which directly train and predict on the original images,avoid the information loss caused by preprocessing,and introduces a novel network structure with better feature extraction capabilities,and a sample expansion strategy is developed.The numerical results has shown that the proposed method significantly improves the recognition accuracy rate compared with the previous solutions,and has a better performance on different receipt formats,and also stronger versatility.Since the name and address information on receipts sometimes contain uncommon characters,this thesis proposes a novel scheme from both the data and algorithm level to improve the recognition performance of rare and low-frequency characters.At the data level,this thesis randomly combines the characters to be recognized into a corpus and generates corresponding training samples to increase the frequency of these characters;at the algorithm level,the proposed method removes the loop layer in the network to reduce the model size and improve training efficiency,and also use different loss functions in different training stages corresponding to the training data.Experimental results show that this scheme can effectively improve the model's recognition performance on rare and low-frequency characters.