Automatic Photo Selection for Media and Entertainment Applications

We propose an algorithm for automatic photo selection for media and entertainment applications like photobook and slide-show. The technique comprises three main steps: photo quality estimation and elimination of poor-quality photos, adaptive quantization of survived photos in time-camera plane, and selection of the most appealing photos from each quantized group. For detection of low-quality photos complex classifier comprising of two AdaBoost classifiers committees is created. Photos with exposure defects, such as overand underexposed, backlit, blurred photos as well as images affected by strong JPEG artifacts are detected confidently. For quantization of photos the method similar to median-cut color quantization is proposed. The appealing photos are selected basing on novel scheme via comparison of visual salience among several images as well as face detection. Our method of identification of the most salient photo among others is based on Itti-Koch-Niebur algorithm of saliency map building. Obtained results of selection as well as time performance issues are discussed. The majority of observers were pleased with the results of the algorithm.