Image restoration and key field alignment for misaligned overlapping text in secondary printing document images.

With the advancement of information technology, the demand for efficient recognition and information extraction from paper documents in industrial scenarios has grown rapidly. In practice, business information is often secondarily printed onto pre-designed templates, which frequently leads to text misalignment or overlap with backgrounds and tables, thereby significantly impairing the accuracy of subsequent Optical Character Recognition (OCR). To address this issue, this paper proposes a preprocessing method for OCR recognition of secondary printed documents, specifically targeting the problems of text misalignment and overlap. In particular, we design a Text Overlap Restoration Network (TORNet) to restore document images affected by text overlap. Experimental results demonstrate that, compared to the latest image restoration models, TORNet achieves PSNR improvements of 0.17 dB and 0.12 dB in foreground and background text restoration, respectively. Furthermore, to resolve residual misalignment issues after image restoration, a key-field alignment method is introduced. This method accurately locates the positional deviations of critical fields in the reconstructed image, enabling precise field-level alignment and structural correction. Based on the proposed preprocessing framework, the recognition accuracy and field-matching accuracy are improved by 23% and 31%, respectively, compared to existing commercial OCR models, significantly enhancing the recognition performance on misaligned and overlapping documents. This study provides an effective solution for recognizing secondary printed documents with text overlap in industrial environments.