Image Scene Classification Based On Feature Learning

With the development of information technology, people pay more attention to images as they can describe and store information quickly and efficiently. It has become a problem in computer vision that how to make computer understand scene categories automatically. Scene classification is a technique that can label an image among a set of semantic categories automatically. It provides contextual information for object recognition. Our paper has three improvements on scene classification:(1) We present a method of scene classification based on the incorporation of a multi-resolution representation into a bag-of-features model. Firstly, we extract scale invariant feature transform from images and use the bag-of-words model to represent images. Then we represent images by Latent Dirichlet Allocation. Finally, we use SVM(Support Vector Machine) classifier to get the scene categories. Our method has a good performance on scene classification by extracting spatial information.(2) We propose a simple but efficient approach by building a multiscale codebook on image features. In this approach, we extract features on images in multiple scales and build a codebook for every scale features. Then we represent images by Latent Dirichlet Allocation in every scale. As it gets more information from images, our approach has a good expansibility to classify scene images and remote images.(3)We improve the classification accuracy by cutting images into patches. By extracting features from these patches, this method can receive more spatial and contextual information. Our algorithm has a good robustness.