Multi-label Scene Classification Based On SIFT Representation And Sparse Coding

As a rapid and effective descriptor and storage medium, the images draw a comprehensive attention with the development of information technology and the practical demands. It is urgent in the computer vision community that leads a computer to acquire the requisite information quickly and accurately from an ocean of images according to the way of human interpretation. Multi-label learning is an effective learning framework proposed for multiple object exists in the real world, which is extensively used in target classification and recognition at present. Scale-invariant feature transform( SIFT) can overcome the classification error resulted from images translation, rotation, brightness, and scale changes, and it is not sensitive to light, noise, microscopic angle change, objects that are partially shaded is quite high. The SIFT algorithm makes great achievements in text classification, nature scene classification and video classification. Based on the SIFT, this paper proposes some multi-label classification methods with the help of sparse coding( SC) and locality-constrained linear coding( LLC) of spatial pyramid model, and also with the help of multi-label k-nearest-neighbor and Rank-SVM.1. A multi-label classification method for natural scene images based on SIFT representation and sparse coding is proposed. Firstly, the images are characterized and encoded by sparse coding of spatial pyramid model under the dense SIFT feature space. Secondly, the images feature distribution is obtained by max pooling. Finally, the classification results of the natural scene images are achieved using two multi-label classification methods, respectively. The proposed method enriches the spatial information of the images and the feature is expressed efficiently, leading to the better classification performance.2. A multi-label classification method for natural scene images based on SIFT representation and locality-constrained linear coding is proposed. In the first place, the multi-scale dictionary is constructed by the dense SIFT features which are extracted from the images in different scales. The images feature is represented and encoded by the established dictionary and the local-constrained linear coding of spatial pyramid model together. Secondly, feature vector statistics is estimated by the max pooling of images feature distribution and the dimension of feature is reduced by restricted Boltzmann machine. Finally, we use two kinds of multi-label classification approaches to recognize the label of natural scene images. The proposed method not only enriches the spatial information of images, but also enriches the scale information of feature. Moreover, it is a rapid and effective way for encoding as the locality-constrained linear coding is. There is one more point I should touch on, that is, the data after the dimensionality reduction by the restricted Boltzmann machine saves the time of learning and classification of our classifiers, and also get higher classification accuracy.