Reach Of Fast Image Retrieval Technology Based On Color Feature

With the rapid development of personal computer, Internet and mobile device, the digital image related applications have enriched our lives. The Content Based Image Retrieval (CBIR) technology has been the hot issue in recent years. This paper will discuss the key techniques in the CBIR system and will also present novel methods in the low-level color feature extraction, resolution of semantic gap and relevance feedback area to meet faster and more accurate retrieval demands.Firstly, we present a new color feature——Color Mutual Information (CMI), which blends mutual information concept in the Color Correlogram (CC). It makes full use of the spatial correlation among different colors in the CC and utilizes the mutual information to compress the feature of CC. In this paper, we combine Color Autocorrelogram (CA) and CMI from CC to generate a new color descriptor——Color Autocorrelogram Mutual Information (CAMI). CAMI not only overcomes the deficiency of the CA algorithms, but also reduces retrieval time and enhances the retrieval precision. It can be used as a real-time retrieval method.In order to reduce the semantic gap and further compress the image feature, we use Support Vector Machine (SVM) and Fuzzy C Means (FCM) methods to classify CAMI feature separately. We obtain the decision value matrix from SVM and degree of membership matrix from FCM. The two features which got from SVM and FCM are called CAMI-SVM and CAMI-FCM. Since SVM and FCM are both methods based on high-level machine learning algorithm, this processing method could reduce the semantic gap between low-level and high-level. We integrate these two features into a new feature——CAMI-SVM-FCM, the size of which is proportional to the number of category in the image database. As for the image database with a relative small number of category, we could use this method to reduce the size of feature vector and faster retrieval speed significantly.Lastly, we present a new relevance feedback method based on category-weight to reduce the semantic gap, which attaches a different "reward" or "punish" constant to the relevant or non-relevant image categories. This method not only improves the retrieval accuracy of CBIR system, but also owns a fast convergence attribute which enables the user get a satisfactory retrieval result by implementing few feedbacks.We build a CBIR system based on the Corel 1K image database to test the CAMI-SVM-FCM feature and category-weight relevance feedback method. The experimental result show that, the CBIR system would get a remarkable improvement on retrieval accuracy by utilizing CAMI-SVM-FCM feature. User could also obtain a satisfactory result after only one relevance feedback.