Abstract: In order to overcome drawbacks of SPM based scene recognition, a novel scene recognition method based on Super-pixel Spatial Pyramid Model（S-SPM） is proposed：images are hierarchically divided into sub-regions through super-
pixel segmentation and then the visual vocabulary is constructed by clustering the Principal component Analysis of Census Transform（PACT） features which are extracted from every corresponding spatial sub-region. To recognize the category of a scene, PACT of all spatial sub-regions are concatenated to form a feature vector and then the bag of words feature are added. Finally the LIBSVM is used to classify the scene category. The experimental results indicate that this method has higher recognition rate.

Key words: scene recognition, S-SPM（Super-pixel Spatial Pyramid Model）, PACT（Principal component Analysis of Census Transform）, bag of words, SVM（Support Vector Machine）

摘要： 针对以往场景识别研究中将图像分割成大小相等的矩形区域进行特征提取而导致识别率低的问题，提出了一种基于超像素空间金字塔模型的场景识别方法：先对图像做不同分辨率的超像素分割，在得到的每个图像子区域中提取PACT特征，然后利用K-means聚类构建出图像集的视觉词典。在进行场景识别时，将每幅图像所有分割子区域的PACT特征连接成一个特征向量，并加入bag of words特征进行分类，最终的场景分类结果在支持向量机LIBSVM上获得。实验结果表明该算法能够有效提高识别率。

关键词: 场景识别, 超像素空间金字塔模型, 空间PACT, bag of words特征, 支持向量机