Abstract:  This paper addresses the problem of low OCR recognition performance caused by character blurring in paper invoices. A novel adaptive iterative visual semantic model is proposed to tackle this issue. The model consists of two modules： the recognition module utilizes ResNet as the encoder and Transformer as the decoder to make initial predictions on the blurred text. The correction module takes the recognition module’s predictions and feeds them into a bidirectional language model， which leverages contextual semantic information to refine characters， thereby performing initial text correction. The results are then input to a discriminator， which outputs them directly if successful or iterates the language model for further refinement if failed， effectively improving the recognition accuracy. Experimental results demonstrate that the proposed model outperforms the current state-of-the-art Chinese recognition model ch_PP-OCRv3 by 3.39 percentage points in recognition accuracy and achieves an average 6.81 percentage points improvement compared to other models. Moreover， the model exhibits excellent generalization performance on public datasets such as IC15， IIIT5K， and IC03-Word， validating its effectiveness.

Key words: text recognition, blurry text, paper invoice, neural network, ResNet