低照度图像文本区域的RTE-UNet++增强算法

RTE-UNet++ enhancement algorithm for text regions in low-light images

摘要:

现有低照度增强算法大多关注图像全局对比度的提升和色彩的保持效果，对于图像中文本区域细节的恢复存在一定不足。为此，构建了具有详细标注信息的多亮度配对低照度文本图像数据集。基于Retinex理论模型和UNet++网络结构，提出了一种面向文本信息增强的RTE-UNet++模型结构。通过设计U型分支结构作为文本区域感知模块和引入文本检测模型，有效提升了模型对文本区域的关注度。同时，在训练阶段采用边缘感知损失进行端到端的优化，提高了字符的对比度和细节信息。实验结果表明，提出的模型在MSRA-TD500数据集和SID数据集上文本区域的图像恢复质量以及文字检测任务性能方面表现出显著优势，验证了该方法的有效性。

Abstract:

Existing low-light enhancement algorithms primarily concentrate on improving the global contrast of images and maintaining color consistency. However, they have certain shortcomings in restoring fine details in text regions. To address this, this paper constructed a novel multi-brightness paired low-light text image dataset with comprehensive annotations. Based on the Retinex theory model and the stucture of UNet++ network, it proposed an RTE-UNet++ model structure for enhancing text information. By designing a U-shaped branch structure as a text region perception module and introducing an integrated text detection model, the proposed model can more effectively enhance text regions. Furthermore, an edge-aware loss function was employed to facilitate end-to-end optimization during the training to improve the contrast while preserving finer details of text characters. Experimental results show that the proposed model significantly outperforms existing methods in terms of image restoration quality and text detection performance on the MSRA-TD500 and SID datasets, validating its effectiveness.