Neural networks and genetic algorithms for learning the scene illumination in color images

This paper proposes the use of neural networks to estimate the scene or camera white balancing illuminant in color images. A real coded genetic algorithm is also used to shirk the size of the neural network input data (thus, a smaller network that will train and run faster), and to identify the areas of a scene that contain the most information about, and therefore best represent the scene illuminant (this offers better data quality thereby improving the accuracy). In addition structural learning with knowledge was incorporated with the back propagation method to further reduce the size of the neural network. The final appearance of the colors in an image depends on many factors. The three main factors are the scene illumination, the illuminant used to white balance the image acquisition device and the properties of the object. While both the camera white balancing illuminant and the properties of the object might be known a prior, it is not always possible to control the scene illumination. In this paper, facial skin color was used as the object interest. Given a visual scene, the proposed method first searches for the presence of at least one face in the scene. If one is found, the genetic algorithm is used to extract the inputs to the neural network, that in turn outputs the estimated scene illuminant. Presently, the proposed method is used to estimated four illuminants in the range 2300K to 6500K.