基于IB-UNet的零参考低照度图像增强方法

A Zero Reference Low Illumination Image Enhancement Method Based on IB-UNet

摘要: 在图像增强领域，成对数据的过分依赖可能导致模型过度拟合，影响其泛化能力。为了解决这一问题，本文提出了一种无监督学习方法，该方法受到Zero-DCE网络架构的启发，采用了基于IB-UNet的零参考图像增强方法。该方法直接学习图像的深层特征，提高了图像特征纹理的提取效率，从而在不依赖成对参考数据的情况下，有效提升低照度图像的增强质量。通过客观指标评估，结合不同模型的对比试验与消融实验，客观验证了所提出模型的优势，展示了其在图像增强任务中的潜力和实用性。

Abstract: In the field of image enhancement, an over-reliance on paired data can lead to overfitting of the model, which affects its generalization ability. To address this issue, this paper proposes an unsupervised learning method inspired by the Zero-DCE network architecture, which employs a zero-reference image enhancement method based on IB-UNet. This method directly learns the deep features of the image, improving the efficiency of extracting image feature textures, thereby effectively enhancing the quality of low-illumination images without relying on paired reference data. Through objective metrics evaluation, combined with comparative experiments of different models and ablation experiments, the advantages of the proposed model are objectively verified, demonstrating its potential and practicality in the task of image enhancement.