基于细节显著性估计的低照度图像增强方法

Low-light Image Enhancement via Detail Saliency Estimation

在基于Retinex分解的低照度图像增强算法中,通过同时估计反射层和亮度层进行反射层恢复是一种流行且有效的方法,但算法结构较为复杂,实现难度大。这是因为恢复反射率是一个约束优化问题,不能用简单的优化技术来解决。本文提出了一种细节显著性估计方法,可以利用简单的梯度下降优化技术从图像中恢复出反射层。该方法是基于我们所提出假设——暗区域近似(dark region approximation,DRA):由于低照度图像暗区域中的光照很弱,因此将其忽略不计,即将输入图像中暗区域的灰度分布直接近似为反射层。首先利用高斯场准则构建目标函数,通过基于DRA的Retinex模型估计细节显著层;然后用拟牛顿法求解该无约束优化问题;最后,从细节显著层中恢复出反射层作为最终增强结果。实验结果表明,与现有同类方法相比,我们的方法在增强效果和计算效率方面都具有优势。

Reflectance recovery via the simultaneous estimation of reflectance and illumination is a prevalent and effective solution for image enhancement based on retinex decomposition,but its use results in a complex algorithm structure because reflectance recovery is formulated as a constrained optimization problem that cannot be solved via simple optimization techniques.In this study,a detailed saliency estimation method is proposed to recover reflectance from grayscale images via optimization employing gradient descent algorithms.This method is built on our hypothesis of dark region approximation(DRA).Because the illumination in dark regions of a low-light image is weak to the point of being negligible,the intensities of dark regions in the captured images are approximated as reflectance.The Gaussian field criterion is applied to establish a differentiable optimization function via DRA.This unconstrained optimization problem is then solved using the quasi-Newton method to estimate the detail saliency layer via the DRA-based retinex model.Finally,the reflectance is recovered from the detailed saliency layer.The results for a variety of images demonstrate the superiority of our method over several state-of-the-art methods in terms of enhancement efficiency and quality.