Automatic image scene detection is very important to image annotation and semantic retrieval. According to the requirement of application, eight specific image scenes such as meeting, mass, beach, etc. were focused on. First, to extract the local features of images, the local key points were detected and reduced, and then the SIFT feature descriptors were calculated. Second, a multi-classifier based on support vector machine was constructed and the features for training were selected to achieve relatively accurate detection results. The experiments were designed to mainly focus on two problems, namely the decision of kernel function of classifier and the strategy of feature selection. Experimental results show that the method can achieve relatively accurate and robust results by using radial basis kernel function to construct classifier and the feature extraction strategy of selecting the top n key points by the scale size order. This method is simple and fast, and can satisfy the actual requirements of application for relatively high precision.