基于Retinex模型的低光照图像增强算法研究

Research on Low-Light Image Enhancement Algorithm Based on Retinex Model

在弱光或者逆光环境下,光学成像设备获取的图像不仅会导致视觉体验不佳,还可能造成可见度低、颜色失真和存在测量噪声的影响,降低视觉系统后续处理的性能。为了提高图像的可见度和视觉系统的性能,本文提出了一种基于Retinex模型的端对端的低光照图像增强算法。该算法改进了基于最大熵的Retinex模型,并采用自适应动态调整曲线增强该模型分解的光照图的对比度,并融合分解出的反射图得到最终的增强图像。该算法是非常轻量级的,训练时间仅需80 s。实验结果表明,该方法与现有代表性方法相比,在视觉效果和常用的四个客观图像评估指标上,均有很强的竞争优势。

Images acquired from optical imaging devices in a low-light or back-lit environment usually lead to poor visual experience. The poor visual visibility and the attendant contrast or color distortion may degrade the performance of the subsequent vision processing. To enhance the visibility of low-light image and mitigate the degradation of vision systems, this paper proposes an end-to-end low-light image enhancement algorithm based on the Retinex model. The proposed algorithm improves the Retinex model based on maximum entropy, and uses adaptive dynamic adjustment curve to enhance the contrast of the illumination map decomposed by the model, and then merges the decomposed reflection map to obtain the final enhanced image. The network architecture of the proposed algorithm is very lightweight, and the training time only needs 80 seconds. Experimental results show that compared with the existing representative methods, this method has a strong competitive advantage in terms of visual effects and four commonly used objective image evaluation indicators.