Feature Extraction And Feature Fusion For Content-Based Image Retrieval

Recently with the rapid development of the Internet, mobile Internet and related technologies, the number of digital images has showed explosive growth, and the demand for the target image from a large collection of images has become increasingly urgent. The most immediate impact of the sharp increase in the number of images is that the task of image retrieval becomes difficult. Therefore, how to quickly and accurately retrieve the target image from the image database has become a hot research topic. The earliest image retrieval system is based on key words, which is matched with the context of the image (such as title, manual annotation) through the user input. These systems generally have some problems, such as the lack of text information, the different meaning of image and text and so on. To overcome these difficulties, the content-based image retrieval system is proposed. CBIR (Content-Based Image Retrieval) can automatically extract the underlying features (such as color, texture, shape, etc.) from the image. In addition, in the field of computer vision, pattern recognition and data mining, researchers often encounter high dimensional feature vectors which have many interrelated or redundant information and even noise. Side effects of these data can significantly increase the time and space requirements, may also lead to over fitting, low efficiency and poor performance in the learning procedure. In practice only part of the feature dimension is important and discriminatory. Feature selection is to solve the above problems, according to a certain evaluation to select the most relevant feature subset, which not only reduces the feature dimension, but also improves the discrimination of the feature.To sum up, this paper focuses on the comprehensive feature extraction, feature fusion and user relevance feedback technology for content-based image retrieval. In terms of feature extraction, we improved the original method of CoLD (Contrast and Luminance Distribution) and proposed CoCD (Contrast and Color Distribution) with a better performance of feature extraction based on CoLD. In contrast with CoLD describing the texture features and brightness characteristics of the image, CoCD introduces the color distribution of HSV color space to replace the original brightness distribution. We studied the importance of contrast distribution and color distribution, and the weight of the similarity measure. The experimental results show that the feature extraction improved is effective through statistical sampling. We also proposed a new feature extraction algorithm based on statistical sampling, named BMVA (Blocked Mean-Variance Algorithm). User feedback technology is commonly used in CBIR systems to improve the retrieval performance of the system. According to the energy attenuation in the process of information transmission, we proposed a new user feedback technique and applied it to the image retrieval system, which improves the retrieval performance of the system. On the basis of feature selection and information fusion theory, we proposed an unsupervised graph theory-based feature fusion algorithm (UGFF) which can synthesize multi-feature-extraction algorithm and extract a feature vector with low dimension and high discrimination. Using the public-open corel5k (5000 pieces) and Wang1K (1000 images) as the test image database, Euclidean distance, cosine distance and effective flow of sorting algorithm EMR (Efficient Manifold Ranking) as the similarity measure and ranking method, and average precision of MAP (Mean Average Precision) as the evaluation standard, we verified the performance of the proposed two kinds of feature extraction algorithm and user correlation feedback technology and unsupervised graph theory-based feature fusion algorithm. Experimental results show that, the proposed two novel feature extraction algorithm can achieve better retrieval performance, the relevance feedback algorithm can significantly improve the performance of retrieval, and unsupervised graph theory-based feature fusion technique not only reduces the dimension of feature vector and enhance the discrimination of the feature vectors.