LK-Net: Efficient Large Kernel ConvNet for Document Enhancement

Qijun Shi, Hongjian Zhan, Yangfu Li, Weijun Zou, Huasheng Li, Umapada Pal, Yue Lu

Abstract: Various types of degradation in document images, such as blurring, shadow, and physical wear and tear, significantly impact the effectiveness of downstream tasks in multimedia applications. The need for document image enhancement arises from the urgent need to improve the legibility and quality of these images, which are integral for accurate Optical Character Recognition(OCR), information retrieval, document analysis, etc. This paper introduces a novel and simple approach employing Large Kernel Convolutional Networks (ConvNets) for document image enhancement, capitalizing on their ability to encapsulate expansive contextual information to improve image quality. Extensive experimental evaluations across multiple benchmarks have demonstrated that our method achieves state-of-the-art (SOTA) while maintaining low computational cost. Code and pre-trained models are available at https://github.com/qijunshi/LKNet.