Scene Classification Based On Single-layer Sparse Auto-encoder And Support Vector Machine

With the coming of “everywhere is picture” era, the demand of automatical image annotation becomes more and more urgent. As the important topic of image classification, the scene classification has attracted great attention. Scene classification is the hot focus of the computer vision. It has many applications, such as content-based image retrieval, object recognition, intelligent video surveillance, digital photo-finishing and mobile robot. Thus far, researchers have achieved encouraging performance in scene classification. However, scene classification remains to be a challenging problem due to its big intra-class variability and small inter-class similarity.This thesis focuses on the problem of scene classification, including extracting the sufficient and appropriate features from the scene images, researching the dimension reduction method for image representation and studying the parameter optimization algorithm for the classifier. The main work and achievements of the thesis can be concluded as follows:Firstly, this thesis aims at extracting the sufficient and appropriate features from the scene images. Firstly, the difficulties and recent methods for scene classification are analyzed. Moreover, the advantages and disadvantages of the recent methods are also analyzed. Secondly, breaking the conventional way of hand-design features, this thesis uses the single-layer SAE to extract sufficient and appropriate features from the scene images without prior knowledge. After extracting patches randomly from scene images,pre-processing is applied to these patches. The single-layer SAE is trained by patches.The image representation is accomplished by this trained single-layer SAE. In this process, the locally connected single-layer SAE network is constructed. Compared with the conventional feature extraction methods, the single-layer SAE can extract more sufficient and appropriate features with low computational cost, which is the main reason of high classification accuracy.Secondly, studying the dimension reduction technique for image representation.The dimension of the image representation obtained by the trained SAE is high. This thesis proposes a dimension reduction method for the image representation based on mean pooling operation, which is a mean filtering technique in essence and can denoise the representation. After mean pooling operation, the feature vectors of the scene images are obtained.Thirdly, researching the parameter optimization algorithm of the SVM classifier.The thesis focuses on studying the parameter optimization algorithm based on PSO.PCA algorithm is applied to extract the main information of the feature vectors and can speed up the process of parameter optimization. The one-versus-one strategy is employed for the multiple scene classification problemFinally, to show the efficiency of the proposed approach, several public data sets are employed. In order to show the proposed method can overcome the color constancy problem, a new data set is constructed. The results reveal that the proposed approach achieves better classification accuracy than the existing state-of-the-art methods.