On over-exposed region detection with regularized logistic regression

Over-exposed region detection is commonly used for image editing and other applications such as over-exposure corrections. In broadcasting, over-exposed regions in image/video can be detected and corrected prior to sending them to consumers. However, the current detection of over-exposed regions is sometimes inaccurate. The detected over-exposed regions are susceptible to noise and scattered. In this paper, we proposed detecting over-exposed regions in images using L2 regularized logistic regression (LR). Over-exposed regions are accurately detected using several novel features in the LR model. These include modeling the characteristics of over-exposed regions as clusters rather than isolated pixels. In addition, overexposed regions are also characterized by both intensity and chrominance pixel values. Optimal parameters for the classifier are obtained by performing training on different scenes. The experimental results show that the detected over-exposed regions are more spatially connected and perceptually accurate compared with current techniques with comparable fast processing.