Research On Image Classification Algorithm Based On Local Feature And Feature Representation

Image classification is a hot issue in the field of computer vision research and plays an important role in the application of image retrieval, image annotation, surveillance video analysis, and so on. In recent years, the development of machine learning, artificial intelligence and multimedia information technology have greatly promoted the research and application of image classification technology. Although image classification has formed a set of mature process, it is still challenging to design classification algorithm with good performance and efficient calculation. Especially, image classification technology is still faced a variety of difficulties in the practical application, such as illumination variation, gesture variation, occlusion, scarcity of training data and differences in sample distribution,etc.In this context, aiming at some problems existing in image classification technology, research is carried out focusing on image representation and classification model by means of related methods and techniques in computer vision and machine learning. The main contents and contributions of this dissertation are briefly stated as follows:1. We propose an accelerated image classification method based on Naive-Bayes Nearest-Neighbor.Naive-Bayes Nearest-Neighbor (NBNN) algorithm effectively avoids the quantization error caused by Bag of Visual Word model (BoVW). However, the method employs the only nearest neighbor for classification and has limitation of slow speed and noise vulnerability. Aiming at the above-mentioned problems, retaining merits of the original NBNN algorithm, an accelerated Naive-Bayes K-Nearest Neighbor classification algorithm is proposed. On the one hand, K-nearest neighbor is employed for classification decision and the influence of background information is also removed. On the other hand, feature selection is included for reducing feature number of test and training images, which improves speed of the algorithm and reduces the impact of noise during classification. And an attempt is also tried to balance the contradictory between classification accuracy and classification time by reducing feature number of test images and training images simultaneously.2. Under the framework of the Naive-Bayes Nearest-Neighbor, we propose an image classification algorithm based on low rank and sparse decomposition and collaborative representation.At present, in order to obtain high classification performance, most of image classification algorithms based on parameter learning need adequate training images. But, in practical applications, issues of scarcity of training samples or high acquisition cost often happen. And noise, illumination, occlusion, complex background and other factors further exacerbate the above problems. On the other hand, although there are some differences between images within the same class, they still share a lot of similarities and correlate with each other, which will be beneficial to the final classification if this property is made full use of. To this end, the paper proposes a non-parametric learning image classification algorithm. Based on non-negative sparse coding combined with max pooling, training images with the same label are expressed as feature matrix with low rank property. On the basis of this, dictionary and low-rank projection matrix are learned by means of low rank and sparse decomposition with structured incoherence constraint. During classification process, the low-rank projection matrix is employed in order to remove the interference information of test image. Then adopting collaboration representation. under the framework of the Nai’ve-Bayes Nearest-Neighbor, the classification process is implemented. Based on the above ideas, an image set classification algorithm is also proposed under the assumptions of acquiring multiple test images with the same class label at once. Finally, the proposed algorithm is experimental analyzed in different standard image datasets for classification.3. We propose a domain adaption image classification method based on multi-sparse representation and online dictionary learning.Generally, research of traditional image classification algorithms assume that training samples and test samples are derived from the same domain (data base) with the same distribution. Unfortunately in practical applications, this assumption is rarely met. According to this problem, domain adaption image classification algorithm based on feature representation is proposed in the dissertation.For reducing the distribution of samples in source and target domains, supposing several intermediate domains are interpolated between the source domain and the target domain. Images are represented as BoVW feature vectors with fixed length in the proposed algorithm and subspaces of intermediate domains are modeled by online dictionary learning, which ensures reconstruction error of samples minimum and smoothness of the transition path. Then, an augmented feature vector encoded by dictionaries corresponding to source domain, intermediate domains and target domain is employed for cross domain recognition. Experimental results verify the effectiveness of the proposed algorithm, which further confirms that image classification methods based on domain adaptation receive better performance when the training samples and test samples are derived from different domains.