Automatic Image Thumbnailing Based on Fast Visual Saliency Detection

Image retargeting has seen many applications in areas such as content adaptation for small displays and thumbnailing for image database browsing. Most retargeting methods, however, are too expensive computationally to achieve fast performance on common desktop systems. This work addresses the problem of fast automatic thumbnailing for image browsing. A simple approach of automatic thresholding saliency maps and cropping using bounding box extraction is presented. Eight of the fastest saliency detectors in the literature and three automatic thresholding methods are assessed using precision, recall, F-score and execution time on the MSRA1K dataset. The results show that the approach is computationally efficient and adequate for fast automatic image thumbnailing. In particular, saliency detection with difference to random color samples (RS) thresholded by Rosin's method achieved the best trade-off between execution time and F-score.