| In the latest years, with the rapid development of the multimedia technology, digital images has gradually replaced the traditional image and become the main carrier of images in the field of visual information. Therefore, how to make a reasonable use of the numerous digital images to provide the needed information to the users has become a serious challenge. In view of such demand, content-based image retrieval emerged and quickly became a technology which attracted much attention in the fields of multimedia information processing. This paper studies on the key problems of the content-based image retrieval technology, manages to improve the performance of image retrieval algorithm, the main contents include the following aspects:1. Since single visual feature cannot fully describe the image content, we present an algorithm using a fusion of multiple features to tackle this problem. Firstly, we construct the Exponent Moments using the R, G, B components of the image respectively as the color feature of the image. Then localized angular phase(LAP) is calculated for each pixel of the image in the HSI space, using the intensity component I, and count the histogram of the localized angular phase of each pixel as the texture feature of the image. Finally, the color and texture features are normalized and assigned with the corresponding weights, after that, the calculation of image similarity method is carried out by using the multiple weighted feature and the retrieval results are sent back to users according to the similarity ranking. The ideal outcome of simulation experiment shows that the content-based image retrieval method using the fusion of multiple features can achieve better retrieval results against single visual feature.2. Since the global features of the image are not well-performed in describing the key content of the image, we present an image retrieval algorithm based on interest points. Firstly, the algorithm uses the chromatic and structural characteristic of the image to construct color invariant characteristics. Then we use the scale-invariant feature detector with error resilience(SIFER) to extract the feature points of the image. Finally, the color histogram, distribution entropy histogram and gradient orientation histogram are applied by using the interest points to represent the features of the image. After that, the similarity between images is calculated with a retrieval result sent back to the user. Simulation results show that, the SIFER operator can find the region of interest(ROI) effectively. The use of feature points in ROI, instead of the global features, in image retrieval can achieve an ideal retrieval results.3. Since there is a semantic gap between low-level features and high-level semantic, we present a color image retrieval algorithm based on relevance feedback. Firstly, we perform the KPCA algorithm to reduce the dimension of the low-level features in order to remove the correlation and noise from the features and additionally decrease the time complexity of the feedback process. Then an improved kernel function N-RBF is developed by embedding the mean value and the mean square difference values of feature attributes in Gaussian RBF kernel function. Finally, optimize the kernel parameter and penalty factor. Simulation results show that the proposed algorithm, which composed of dimension reduction, kernel function modification and parameter optimization can effectively decrease the time complexity of the algorithm, meanwhile achieve a comparatively ideal retrieval result. |
