The Research Of Image Scene Classification Algorithm Based On Multi-Level Feature Representation

Image scene classification has become one of the most common image recognition tasks among computer vision research area, which aims to extract and analyze image features and assign various images to the same category sharing similar contents. As for computer, the input of images is in the form of digital matrix. Given huge digital information, extracting useful representations to distinguish scene classes has become a gap between the “digital storing similarities” of computer and “concept and content similarities” of human.Therefore, computer tries to simulate the principle of human convex recognition process when dealing with scene image tasks, which involves multi-level features. It includes the low-level features, generating from pixel of image; the middle-level features based on quantizing and encoding method; high-level features, such as object attribute based sematic features. Among the three level based features, the low-level feature is the basis of the image understanding, the Object Bank feature based on object attribute is characterized by semantic objects in the scene image, while the middle-level feature aims to connect both the two features. Building multi-level feature representation enables computer to cover the associations of different level based feature. Moreover, multi-level representations are complementary and reinforcing, which is better to explain the content of images and identify the scene categories efficiently. In this way, computer may recognize scene images like human brains. The main contents of this paper are as follows:(1)We proposed a fast Locality-constrained Low-rank coding algorithm for image classification. The method encodes the SIFT descriptors of low-level features jointly. Simultaneously, we added locality-constraint regularization, describing the correlations of the local features. Furthermore, we replaced the nuclear norm with the Frobenius norm for approximate optimization method, resulting in small computation and complexity. Specifically, firstly, our method applied the fast Locality-constrained Low-rank coding algorithm to SIFT descriptors, obtaining the coding representations; thus, pooling and Spatial Pyramid Matching methods were added to get the ultimate representation matrix; finally, we learnt a SVM classifier according to the training sample labels and representations, achieving the classification of unknown images.(2)We proposed a Cascaded Representation based on multi-level features algorithm for scene image classification. Middle-level feature representations based on Low Rank coding encode and quantize the low-level descriptors, describing the correlations of features; in addition, images often display overall layout relationships between the different areas, while these features are characterized by objects attributes of high-level features. Therefore, in the first place, we use Object Bank representation to describe that object semantic features, learning SVM classification model of scene topics. So it obtained a crude understanding of the scene image, resulting that scene classes containing similar object features are assigned to the same scene theme. Afterwards, as for the same scene topic may include many different scene classes, based on the Locality-Constrained Low-rank coding method mention above, we applied it to SIFT descriptors, encoded the low-level features to middle-level features. It then learnt a single SVM classifier to distinguish the scene classes within the scene themes in detail. In the way of cascaded representation ranging from middle-level features to high-level features, it achieved the rough understanding based on object semantic features to detailed analysis based on middle-level features progressively and complementarily.In this paper, we use Matlab simulation software and LibSVM toolkit, simulating and doing experiments on the traditional scene classification database. We verified the efficiency of our proposed algorithms from average classification accuracy and computing time two aspects. Experimental results show that fast Locality-constrained Low-rank coding algorithm significantly reduced the computing cost compared to the traditional algorithm, additionally, the classification result of Cascaded Representation based on multi-level features algorithm has a further improvement.