Using convolutional neural network models illumination estimation according to light colors

Abstract

One of the important problems in digital images is that the object's color changes as the light source's color changes. Color constancy methods are used to solve these problems. Color con-stancy is the ability to detect object color in an image under lighting sources accurately. The light color of the image is estimated by calculating color constancy. Many statistical and learning-based approaches related to calculating color constancy have been presented. In recent years, deep learning algorithms, one of the learning-based approaches, have been used to calculate color constancy. This study aims to estimate the illumination of images obtained in varying light colors with deep learning methods. This study was applied to the agricultural sector. The illumination estimation was performed using the 3-fold cross-validation method with VGG16, EfficientNet-B0, ResNet50, MobileNet, DenseNet121, and GoogLeNet models. A transfer learning approach was adopted in this study. Illumination estimation was applied to a new data set. The median value of the angular error (AE) metric performed well in all experimental results. The lowest AE values were obtained in the proposed GoogLeNet model. This model AE values: the mean was 2.220, the median was 2.126, the trimean was 2.006, and the maximum was 6.596 degrees. In addition, the number of images with AEs below 3.0 constituted 77.13% of all images. The results of the Friedman and Wilcoxon signed rank tests confirmed the effectiveness of the proposed GoogLeNet model in illumination estimation.