Abstract: An image recognizing method adopted by a platform is disclosed. The method first receives multiple targets to be recognized at the platform, and inquiries a pre-established semantic tree by reference to the targets for determining if the recognition results of the multiple targets will cause confusion or not. If confusion is not foreseeable, the method obtains respectively a parent-classifier corresponding to each parent-category of each of the targets, and uses the parent-classifiers directly to perform a recognition action to the targets. Otherwise, the method obtains respectively multiple child-classifiers corresponding to multiple subcategories below each of the targets, and uses the multiple child-classifiers to perform such recognition action to the targets.

Abstract: A computer-vision based automatic inserting and playing method for advertisement analyzes a video for obtaining multiple shots of the video, and obtaining semantic information related to targets appearing in each shot and low-level information related to vision of each shot. Next, the method matches the semantic information and the low-level information of each two adjacent shots with other semantic information and other low-level information of candidate advertisements, so as to determine one or more insertion time points of the video, and one or more inserting advertisements respectively related to each of the insertion time points. The matching of the semantic information ensures the inserting advertisement and a front shot and a rear shot of the insertion time point are all related with each other, and the matching of the low-level information enhances visual fluency in playing and switching among the front shot, the inserting advertisement and the rear shot.