Iterative Convolutional Neural Network-Based Illumination Estimation

In the image processing pipelines of digital cameras, one first steps is to achieve invariance in terms scene illumination, namely computational color constancy. Usually, this done two successive which are illumination estimation and chromatic adaptation. The aims at estimating a three-dimensional vector from pixels. This represents it used adaptation step, eliminating bias colors caused by illumination. An accurate crucial for successful However, an ill-posed problem, many methods try comprehend with different assumptions. paper, iterative method proposed. calculates series intermediate estimations adaptations input using convolutional neural network. network has been trained iteratively compute incremental estimates original image. Incremental combined per element multiplication obtain final estimation. approach aimed reduce large errors usually occurring highly saturated light sources. Experimental results show that proposed outperforms vast majority median angular error. Moreover, worst-performing samples, i.e., samples errs most, all other margin more than 18% respect mean third quartile.