CG-DIQA: No-reference Document Image Quality Assessment Based on Character Gradient

Abstract: Document image quality assessment (DIQA) is an important and challenging problem in real applications. In order to predict the quality scores of document images, this paper proposes a novel no-reference DIQA method based on character gradient, where the OCR accuracy is used as a ground-truth quality metric. Character gradient is computed on character patches detected with the maximally stable extremal regions (MSER) based method. Character patches are essentially significant to character recognition and therefore suitable for use in estimating document image quality. Experiments on a benchmark dataset show that the proposed method outperforms the state-of-the-art methods in estimating the quality score of document images.

• The paper introduces CG-DIQA, which utilizes character gradients to assess document image quality without a reference.
• It applies the MSER algorithm to extract character patches and the Sobel filter to compute gradients that correlate strongly with OCR accuracy.
• Experimental results show high median LCC (0.9841) and SROCC (0.9429), outperforming traditional metric-based and learning-based methods.

No-Reference Document Image Quality Assessment Based on Character Gradient

Introduction

The paper "CG-DIQA: No-reference Document Image Quality Assessment Based on Character Gradient" (1807.04047) presents a novel approach to document image quality assessment (DIQA), an essential task for ensuring high OCR performance in practical applications. The proposed method, termed CG-DIQA, leverages character gradient information extracted from document images using a no-reference approach. This method utilizes the OCR accuracy as a ground truth quality metric to predict document image quality scores.

Motivation and Background

As the demand for digitizing documents via mobile devices grows, the associated challenges of image degradations—such as blur—become more pronounced, adversely affecting OCR systems. Traditional natural image quality assessment methods fail to adequately address these document-specific challenges, necessitating the development of specialized DIQA techniques. Two predominant DIQA approaches exist: learning-based methods, which require extensive datasets, and metric-based methods, which depend on handcrafted features often inefficiently extracted from non-character-centric patches. This paper seeks to address these limitations using character gradient information for quality prediction.

Methodology

The CG-DIQA method consists of several key steps:

1. Preprocessing: Conversion of the input document image to grayscale followed by downsampling to manage computational resource allocation efficiently.
2. Character Patch Extraction: Utilizing the MSER algorithm, the method extracts patches within the document that are likely to contain characters. This approach prioritizes regions crucial for OCR, thus optimizing the patch extraction process.
3. Character Gradient Calculation: The paper identifies the smoothed gradients associated with degraded character edges. These gradients are computed using the Sobel filter to achieve high correlation rates with perceived document image quality.
4. Quality Score Estimation: The overall document image quality is assessed by calculating the standard deviation of these character gradients. Pooling strategies like average pooling are employed to derive final quality scores without incurring significant computational overhead.

Experimental Results

The method's efficacy was tested on a benchmark DIQA dataset comprising 175 document images with varied degradation levels. Compared to existing methods such as metric-based and learning-based approaches, CG-DIQA demonstrated superior performance. The method achieved median LCCs and SROCCs of 0.9841 and 0.9429, respectively, outperforming state-of-the-art NR DIQA methods in both document-wise and overall evaluation protocols.

Implications and Future Directions

The introduction of CG-DIQA highlights the potential of employing character-specific features for document image quality assessment. By focusing on the character gradient, this approach optimizes the accuracy of quality predictions, crucial for enhancing OCR processes in mobile-captured documents.

Future research could explore the integration of advanced feature extraction techniques and adaptive weighting strategies within the CG-DIQA framework. Such enhancements could further refine document quality assessments, providing a robust tool for real-world applications.

Conclusion

The CG-DIQA method presents a significant advancement in document image quality assessment by effectively using character gradients to predict quality scores. The method's reliance on character-centric patches aligns closely with the practical needs of OCR applications, offering a compelling solution to the limitations of existing DIQA approaches. Further research into feature extraction and weighting strategies could yield additional improvements, cementing CG-DIQA's role as a critical tool in the document processing pipeline.