| Summary: | 碩士 === 國立臺中科技大學 === 資訊工程系碩士班 === 101 === With the advances in various multimedia technologies, an explosive growth of multimedia databases and digital libraries are being produced constantly. These large collections of images make difficult for users to search and browse efficiently through the entire database. Therefore, how to get the specific images from the databases is a crucial problem. In this thesis, there are two retrieval methods proposed to deal with this problem.
A directional local pattern-based variability histogram and particle swarm optimization (PSO) is proposed. This method extends the standard local binary pattern (LBP) and local ternary pattern (LTP) by using the directional relationship between the central pixel and its neighbors. Directional local pattern is a structure to encode directional features based on local variations. Firstly, the image is computed to obtain the different values of each pixel with its neighbors in each direction of 0°, 45° and 90° respectively. Then, collect gray-level differences and use PSO algorithm to compute the quantized thresholds. After the above process, construct the variability histogram for the directional local patterns to present the texture feature. Finally, we calculate the similarity distance between query image and database images. And the expected experimental results indicate that our algorithm performs much better than traditional LBP and LTP do in terms of average retrieval rate (ARR).
The multi-element descriptor of texture construction features for content-based image retrieval (CBIR) is the second proposed method. This method extracts color feature and texture features by multi-element descriptor which is constructed by ten element types. Each type denotes individual directional state of 2×2 grid. Firstly, the color image is converted into HSV color space. Then, extract primary color feature from quantized Hue and Saturation components and use multi-element descriptor to extract texture feature from quantized Value component. If extracted texture feature is conformed to the regional threshold, then the color feature of corresponding position can be saved. Finally, combine color feature with texture feature and compute the similarity distance between query image and database images. By the experiments, the proposed method is more efficient than the structure elements descriptor and micro-structure descriptor. The proposed method is tested to Corel-1,000 database and Corel-10,000 database. The experimental results indicate that our algorithm has much better and effective performance in terms of average precision and average recall.
|
|---|
