The invention relates to a method of locating an image foreground using LLC (Locality-constrained Linear Coding) criterion. A large number of random images are selected from a standard test set, in combination of the salient region truth value annotation graph, the priori knowledge of the image foreground is extracted, an LLC codebook is formed, the LLC criterion is used for carrying out rough classification on whether each area of a to-be-detected image belongs to the foreground, and a corresponding salience probability value is given; contrast-based features such as a centroid distance away from an image center, a local Lab color contrast value and a global Lab color contrast value are used for describing image super pixel regions, typical features for learning the foreground/background serve as the priori knowledge for guiding classification of the image super pixel regions, high-level knowledge is acquired from an empirical perspective, region classification can be guided for multiple times as long as one-time learning is needed, the foreground locating speed is greatly quickened compared with a method of extracting high-level knowledge only from the current image, and due to advantage query for extraction, the foreground boundary in the acquired salience map based on manifold ranking can be more clear and has less noise.