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.