Intelligent method for contrast enhancement in digital video

In this paper, an intelligent method for image contrast enhancement in real time digital video applications is proposed. This new technique is based on the generation of adaptive transfer functions by using Neural Network. The method is designed in order to be implemented in digital television systems, based on Thin Film Transistor Liquid Cristal Displays (TFT-LCD), that has become the dominant technology in consumer electronics. The method provides the required amount of contrast enhancement for every image according to real time analysis of brightness and contrast in the video frame information, then a different contrast enhancement transfer curve for each image is generated. The statistical information is extracted by histogram analysis from current or previous image frames. Technological factors are considered where it is applied, as well as aspects of the visual perception within a non-controlled environment, which is the case of the consumer environment. The method is based on the design of a general transfer function model that compensates the decrement in gray scale dynamic range in images representation which is a consequence of the TFT-LCD technology limits; in particular the light leakage from the panel backlight sources of the Cold Cathode Fluorescent Light (CCFL) or the LED type, as well as an unsuitable dynamic control of backlight intensity. Using subjective methods for image quality, the contrast is enhanced in a set of representative images for different degrees of brightness and contrast, a unique transfer function is obtained for each representative image. This set of transfer functions, as well as the statistic of the pixel values distribution from each image frame are used as an input - output pattern during the training of a neural network. On the basis of the experimentation, an evaluation and comparative analysis takes place between the functions of transformation obtained by subjective method of image quality and the curves obtained through the trained neural network.