Analysis of the effect of lighting on image processing problems using artificial intelligence techniques

The appearance of an object is affected by the color and quality of the light falling on the object's surface and the location of the illumination source. These situations can make object classification and recognition difficult in machine vision applications. For the quality of the images collected using different lighting strategies during the image acquisition stage differs from each other. Therefore, the change in lighting conditions causes the differentiation of the information obtained from the images by image processing techniques and limits the performance of the algorithms used to make sense of it. Convolutional neural network (CNN), one of the deep learning architectures in recent years, has been used in image classification processes to solve this problem. Within the scope of the thesis, CNN-based studies have been proposed for the classification of images affected by lighting differences and the estimation of illumination and luminance. This thesis study has been applied to physiological disorders in the apple sector. This study tried two lighting scenarios depending on the light colors. According to the first lighting scenario, apple images were obtained at different lighting conditions (different light colors and lamp brightness values), position angles, and distances. These images were labeled according to both physiological disorders types and light colors. According to the types of physiological disorders, the original data set consists of 1080 images, and the augmentation data set consists of 4320 images. These data sets were named physiological disorders in apple-1 (PDA-1) and physiological disorders in apple-32 (PDA-32), respectively, and these data sets are in three classes. Also, the data set, separated according to light colors (warm white, cold white, and green light), has a three-class structure. In another scenario, apple images were obtained at different color temperatures of white light (warm, medium, and cool white), illumination position angles, position angles, and distances. These images are labeled according to the defect condition in the apple. This data set consists of 1296 images. Five approaches were proposed using these scenarios in this study. In the first approach, PDA-1 data set physiological disorders images were classified by end-to-end training of CNN models. This approach obtained in the best classification performance in the Xception model. Average accuracy, precision, recall, F1-score, and AUC values of the Xception model were 0.996, 0.994, 0.998, 0.996, and 1.000, respectively. The second approach evaluated hybrid methods that classify physiological disorders with machine learning methods using pre-trained CNN models. This approach has been applied to the PDA-1 and PDA-32 data sets. In both data sets, the highest average classification accuracy was found in VGG19(fc6) and support vector machines (SVM) models with a rate of 0.961. Here, 4096-dimensional deep features were used. In addition, a hybrid study was applied by feature selection in this approach. According to feature selection, the highest classification accuracy of 0.948 was obtained using VGG19(fc6) and SVM model in the 512-dimensional deep features. Another approach focused on the effect of classification performance of images produced by color balancing models (sharpness, gamma correction, and Contrast Limited Adaptive Histogram Equalization) to solve the problems caused by light variation. These data sets were separated according to the light colors in this approach, and the transfer learning approach was used. The highest classification accuracy was obtained in the Xception model with a ratio of 0.934 in the cold white light color, and gamma correction data set type. In addition, it was determined that the data sets created with color balancing improved in Peak Signal-to-Noise Ratio (PSNR) metric. Another approach estimated the illumination of images obtained in different light colors with CNN models. The transfer learning approach was adopted in the study. In addition, statistical and learning-based methods and proposed CNN models were compared. The best angular error (AE) values were obtained in the proposed GoogLeNet model. This model's AE values: the mean was 2.220, the median was 2.126, the trimean was 2.006, and the maximum was 6.596 degrees. The mean absolute percent error (MAPE) of the proposed GoogLeNet model was found to be 14.732%, and it can be said that the illumination estimation of this model is good. In addition, the number of images with AEs below 3˚ constituted 77.13% of all images. In addition, this model determined that the sample images were improved according to PSNR and Blind/No Reference Image Spatial Quality Evaluator metrics. Finally, image-based luminance estimation was made with the help of CNN models using images containing different color temperatures and lighting source position angles. This approach obtained the best luminance estimation in the GoogLeNet model. According to color temperatures, the lowest Root Mean Square Error (RMSE) value was 5.023 cd/m2 in cold white light type. According to the MAPE value, it was determined that the prediction of the GoogLeNet model was good in this light type. The lowest RMSE value was calculated as 5.106 cd/m2 at 60 degrees, according to the lighting source position angle. Here the MAPE value was the lowest percentage. Since the images in the resulting data sets vary according to the light colors and lighting locations, the defect information obtained from the apples are also quite different. According to the experimental results, the classification application's performance decreased with the addition of noise and brightness values to the images.