Scene Classification Via Sparse Representation

In many computer vision applications, image classification is a critical technology. Scene categories technique is more challenging. We usually treat image classification task as object recognition. By telling the object in the image we then give a label for the image. But in the scene images, there is more than one object. Such as in the street image, there are trees, buildings and road and in a countryside image, there are also trees. Therefore, we can’t tell the category only by recognizing an object.When we see a picture, only a part neuron in the receptive field can make a feedback. The theory of sparse representation is similar with this. By simulating human’s vision system, sparse representation is widely used in image processing. This success is mainly due to the fact that images have naturally sparse representation with respect to some fixed bases. Therefore, spare coding is a nice code for images.In this paper, we present a novel scene image classification method which combined with sparse representation. In our method, the image codes are discriminative with other categories but analogous with intra-class.Our method applies a coding scheme forcing Non-negative constraint on sparse coding along with max pooling and spatial pyramid matching method to represent images. To favor the non-negative constraint, the codebook is trained by a large scalable online dictionary learning method. The non- negative sparse coding efficiently reduces the information loss and eliminates ambiguous representations, and thus nicely upgrades the discrimination of sparse codes. The experiment results demonstrate the proposed method outperforms many recently proposed sparse coding approaches for image classification.