基于卷积神经网络的逆光图像增强研究

Research on Backlight Image Enhancement Based on Convolutional Neural Network

现有的大部分算法只能针对特定照度的逆光图像有出色的增强效果,不能高效率地解决各类照度的逆光图像。因此,本文提出一种基于卷积神经网络的图像增强算法,并构建一种集分解、恢复、调节为一体的新型网络架构。利用Retinex理论设计一个分解网络,将逆光图像与其对应的高光图像都分解为反射图和光照图。采用高光图反射分量作为去噪参考,修复暗光缺陷,并添加颜色饱和度模块,最大程度地保留图像恢复过程中的颜色等细节。逆光图像的光照图可依据用户喜好自适应调节亮度,设置增强比率(目标光源与图像光源间的比值)作为调节指标,将逆光图像增强至高光图像时,增强比率要大于1。在多个公开数据集(LOL、DICM、NPE)上验证,研究表明本文方法可有效增强逆光图像亮度,改善图像质量,保证图像细节不丢失,避免颜色失真。在不同照度的逆光图像上均有较好的效果,主观和客观评价指标上的结果优于对比算法,对智慧城市的安防以及人工智能的发展有应用价值。

Most of the existing algorithms can only enhance the backlight images with specific illumination,but cannot solve the backlight images with various illuminations efficiently.Therefore,an image enhancement algorithm based on convolutional neural network is proposed in this paper,and a new network architecture that integrates decomposition,recovery and adjustment is built at the same time.Using Retinex theory,a decomposition network is designed to decompose the backlight image and its corresponding highlight image into reflectance map and illumination map.The reflectance component of highlight image is used as the denoising reference to repair the dark light defect,and the color saturation module is added to retain the color and other details in the image restoration process.The brightness of the backlight images can be adjusted adaptively according to the user’s preference.The enhancement ratio(the ratio between the target light source and the image light source)is set as the adjustment index.When the backlight images are enhanced to the high-light images,the enhancement ratio should be greater than 1.Validated on multiple public datasets(LOL,DICM,NPE),the research shows that this method can effectively enhance the brightness of backlight images,improve image quality,ensure that image details are not lost,and avoid color distortion.It has good effects on backlight images with different illuminations,and the results of subjective and objective evaluation indicators are better than the existing algorithms,which has application value for the development of smart city security and artificial intelligence.