High Dynamic Range Image Generation Based on Convolutional Weight Optimization Robust to Mixed Noise Removal

Abstract

Exposure blending generates a high dynamic range (HDR) image from multiple exposure images. However, when photographing in a low-light scene, these images are deteriorated due to various types of noise, and then it brings down the dynamic range. Existing methods based on pixel-wise blending cannot sufficiently reduce noise, especially in the case of a few input images. This report proposes a convolutional weight optimization method for exposure blending that robustly removes mixed noise and under/over-exposed pixels with a few inputs. In blending each pixel, the proposed method convolves neighboring pixels and generates a noise-free HDR image. The convolution of local regions enables to enhance the denoising capability of image blending. To find a set of weight maps for convolution, we introduce a weight optimization problem as a convex optimization problem, in which Huber loss function is utilized as a fidelity measure in blending to make the method robust to outliers, and solve the optimization problem by using the primal-dual splitting method. The weighted sum of the noisy input images with the estimated weight maps makes a noise-free HDR image. Experimental results show the validity of the proposed method compared with several conventional methods.