Scene illumination classification based on histogram quartering of CIE-Y component

Despite the rapidly expanding research into various aspects of illumination estimation methods, there are limited number of studies addressing illumination classification for different purposes. The increasing demand for color constancy process, wide application of it and high dependency of color constancy to illumination estimation makes this research topic challenging. Definitely, an accurate estimation of illumination in the image will provide a better platform for doing correction and finally will lead in better color constancy performance. The main purpose of any illumination estimation algorithm from any type and class is to estimate an accurate number as illumination. In scene illumination estimation dealing with large range of illumination and small variation of it is critical. Those algorithms which performed estimation carrying out lots of calculation that leads in expensive methods in terms of computing resources. There are several technical limitations in estimating an accurate number as illumination. In addition using light temperature in all previous studies leads to have complicated and computationally expensive methods. On the other hand classification is appropriate for applications like photography when most of the images have been captured in a small set of illuminants like scene illuminant. This study aims to develop a framework of image illumination classifier that is capable of classifying images under different illumination levels with an acceptable accuracy. The method will be tested on real scene images captured with illumination level is measured. This method is a combination of physic based methods and data driven (statistical) methods that categorize the images based on statistical features extracted from illumination histogram of image. The result of categorization will be validated using inherent illumination data of scene. Applying the improving algorithm for characterizing histograms (histogram quartering) handed out the advantages of high accuracy. A trained neural network which is the parameters are tuned for this specific application has taken into account in order to sort out the image into predefined groups. Finally, for performance and accuracy evaluation misclassification error percentages, Mean Square Error (MSE), regression analysis and response time are used. This developed method finally will result in a high accuracy and straightforward classification system especially for illumination concept. The results of this study strongly demonstrate that light intensity with the help of a perfectly tuned neural network can be used as the light property to establish a scene illumination classification system.

Preview Text FK 2014 129IR.pdf Download (1MB) | Preview

Actions (login required)

View Item