Provided is a new learning-based method according to which it is possible to accurately detect a salient region from one image without requiring exemplary data for learning. With respect to an input image, (1) first, a rough foreground map is calculated using a model-based algorithm. (2) Positive learning samples and negative learning samples are respectively extracted from a foreground region and a background region in the rough foreground map. (3) Multiple classifiers are learned using the learning samples defined in step (2). (4) A strong classifier is constructed by combining the multiple learned classifiers, and a final saliency map for the input image is generated using the strong classifier.