Color Constancy Algorithm Based On BP Neural Network

With the continuous development of the information science and technology, digital image technology has been widely applied in many fields. And it has shown great potential. In particular, color image processing technology has become a research hotspot in the field of machine vision. And in machine vision, object recognition has been widely used in many places, such as security surveillance, video content-based image retrieval. One of the most important problems of machine vision is to resolve the color recognition. For the machine vision system, color is considered as an important feature of an object in image understanding, object recognition, etc, and it has played an important role in machine vision systems. When the machine vision system processes color information, it needs the stable description of spectral reflectance characteristics, which relates to the color constancy problem.The ability to compute color constant descriptors of objects in view irrespective of the light illuminating scene is called color constancy. Accurate color reproduction is highly important for many computer vision tasks such as background subtraction technology. To solve the problem of color constancy in the field of machine vision, this paper proposed a new method to update background image based on neural network because of the generalization capability of neural network. By means of appropriate sample sets, we adopt improved back-propagation learning algorithm to train the neural network to obtain the mapping relation of the corresponding pixels of the image before and after the changes. After training, the neural network would output image data with color constancy. The method does not need to build the restricted adaptive model and does not need specific surface property assumptions for the input data. All the rules of color constancy are acquired in the process of neural network training. So the model has a certain degree of robustness, self-adaption, and self-learning. The main innovation of this thesis can be summarized as follows:First, we convert the traditional color constancy problem to the color constancy based on brightness constancy and the color constancy based on color information by Luv color space. And this paper also assumed that the mechanism of color constancy is the result of local scale information exchange determined by the Luv three-channel processor subsystem of the visual system.At last, the method was tested by the experiment of background update under the scenes with randomly selected illuminants, and it was proved to be effective.