应用雾天退化模型的低照度图像增强

发布时间：2018-05-20 17:02

本文选题：低照度图像增强 + 大气散射模型　；参考：《中国图象图形学报》2017年09期

【摘要】：目的针对目前低照度图像增强算法存在噪声敏感、易饱和等现象,提出了一种基于雾天退化模型的低照度图像间接增强算法。方法首先将低照度图像反转成为拟雾图;拟雾图与真实雾天图像有所不同,一是通常具有大面积明亮区域,二是大气光值较高。对于大面积明亮区域,暗原色先验理论并不适用,不容易精确估计相应透射率,因此,提出利用卷积神经网络求解透射率的方法;又针对全局大气光值易出现饱和现象,提出使用局部大气光值代替全局大气光值,从而得到大气光图;之后,利用导向滤波对透射率图和大气光图进行修正;最后基于大气散射模型还原出无雾图像,再次反转无雾图像得到低照度图像的增强结果。结果设计了3组实验,第1组实验为各算法的主观对照,第2组实验从客观指标上对各算法进行比较分析,第3组为实验透射率与大气光值的组合对照实验。结果表明,无论是与Retinex和MSRCR(multi-scale retinex with color restoration)为代表的直接增强算法比较,还是与基于He算法的间接增强等算法相比,本文算法在平均梯度、信息熵、峰值信噪比上均表现良好,且本文算法峰值信噪比平均比次优结果高了2.6 d B,相对应的方差较小,可以有效提高视觉效果,不仅有效提升了低照度图像的亮度,又避免了明显的颜色失真、曝光过度等现象。结论通过定性及定量的实验结果表明,本文算法不仅提高了视觉效果,且场景适应能力较强,能很好地增强室内和室外的低照度图像,且本文算法运行时间中等,若结合cuda技术,还可用于监控视频的实时增强。
[Abstract]:Aim to solve the problem of noise sensitivity and saturation in low illumination image enhancement algorithms, an indirect enhancement algorithm for low illumination image based on fog degradation model is proposed. Methods at first, the low-illuminance image is reversed into a fog-like image, which is different from the real fog image. One is that there is usually a large area of bright area, the other is that the atmospheric light value is relatively high. For a large area of bright areas, the prior theory of dark color priori is not applicable and it is not easy to estimate the corresponding transmittance accurately. Therefore, a convolution neural network method is proposed to solve the transmittance, and the global atmospheric light value is easily saturated. The local atmospheric light value is used instead of the global atmospheric light value to get the atmospheric light map. After that, the transmittance map and atmospheric light map are corrected by the guidance filter. Finally, the fog free image is reduced based on the atmospheric scattering model. The enhancement result of low illuminance image is obtained by inversion of fog-free image again. Results three groups of experiments were designed, the first group was the subjective control of each algorithm, the second group was the comparative analysis of each algorithm from the objective index, the third group was the combination of experimental transmittance and atmospheric light value. The results show that, compared with the direct enhancement algorithm represented by Retinex and MSRCR(multi-scale retinex with color restoration), or the indirect enhancement algorithm based on he algorithm, this algorithm performs well in average gradient, information entropy and peak signal-to-noise ratio. The average PSNR of this algorithm is 2.6 dB higher than the sub-optimal result, and the corresponding variance is smaller, which can effectively improve the visual effect, not only enhance the brightness of the low-illuminance image, but also avoid the obvious color distortion. Overexposure, etc Conclusion the qualitative and quantitative experimental results show that the proposed algorithm not only improves the visual effect, but also has a strong adaptability to the scene, and can enhance the low illuminance images both indoors and outdoors, and the running time of the algorithm is moderate. If combined with cuda technology, it can also be used to monitor the real-time video enhancement.
【作者单位】：福州大学数学与计算机科学学院;
【基金】：福建省产学合作重大项目(2016H6010) 福建省自然科学基金项目(2015J01420) 福建省引导性基金项目(2016Y0060) 福建省卫生教育联合攻关计划项目(WKJ2016-2-26)~~
【分类号】：TP391.41

【相似文献】