Image Retrieval Based On Edge Orientation Difference Feature

With the rapid development of internet technology and information technology, the multimedia information system has been applied extensively beyond the traditional database systems. The application of multimedia information needs a technology to find the image quickly and accurately, which called image retrieval. Therefore, how to retrieve fast and accurately based on visual features of images becomes an important issue in image information filed. This paper proposes a novel method to predict the edge orientation of edge pixels and one new edge orientation difference feature descriptor, and then proposes an effective weighted coding method applied into the image retrieval.(1) This paper proposes one new and effective image retrieval method which named edge-based orientation difference histogram. We firstly obtain the main orientation of each edge pixel through edge direction and energy prediction, and then combine the oriented pixel histogram and edge orientation different histogram to get the edge-based orientation different histogram. We calculate similar measure by rotating the oriented pixel histogram. This proposed method can accurately extract the edge orientation of edge pixels, and this feature is more typical by rotating orientation to improve the performance of image retrieval.(2) An image retrieval method based on the superpixel and image coding is proposed. First, the image is divided into superpixel. And then we calculate the edge orientation different descriptors and scale-invariant feature transform (SIFT) descriptors of dense pixels in each superpixel, and these two descriptors are integrated to present every superpixel. Last, we conduct the vector quantization (VQ) coding for image retrieval using the superpixel as the basic unit. This method can improve the rate of image coding and the efficiency of the image retrieval.(3) An image retrieval method based on weighted coding using edge orientation descriptors is presented. First, we extract the edge orientation different descriptors of edge pixels, and calculate the Color-SIFT descriptors of key points in image. Then, the image coding is achieved based on the two descriptors. Finally, we calculate weighted similar measure and accomplish the image retrieval. This method can update the weights according different retrieval image, and improve the image retrieval precision.This work is supported by National Natural Science Foundation of China (No.60972148).