An Image Retrieval System Based On Binary Descriptors

Image retrieval has become a very popular filed of modern computer areas which is very worthy of study. Finding the query image from massive image data is not only a demand for a for a digital society where people are very anxious, but also is the first step of the machine try to perceive the real world and achieve the true artificial intelligence. Now we have a lot of very mature commercial image retrieval system such as Google and Baidu search map function, and we can also use the SIFT retrieval based on BOW architecture which can be achieved relatively good image retrieval in the laboratory. When we want to combine the image retrieval function with intelligent mobile devices, we will be faces with many problems, including network transmission, database storage, speed of matching calculation and so on. To solve these problems, we hope to improve the image feature extraction, using binary descriptors which are more simply and more efficient on feature extraction and matching. This paper proposed an image retrieval system based on improved binary descriptor called color-orb, and this image retrieval architecture is more accurate and more efficient.In terms of image feature descriptors, different from the general ways of using sift or surf, we chose binary descriptors which are more simple and more efficient matching. On the basis of traditional ORB descriptors, this paper introduced the four-channels color space to improve the color changes and lighting variations robustness of the descriptor. We encoded the gradient information of the three color channels around the key points, and add the color binary code to the traditional ORB codes. It can be seen through the experiments that the color-orb has a better matching performance under various changes and the computation time is almost the same as the traditional descriptor.This paper proposed an image retrieval structure based on the color-orb descriptor. In the first step of the image retrieval process we introduced perceptual hash(phash) algorithm for image database pretreatment. Perceptual hash algorithm can convert an original image quickly to 64-bit binary code, and this code can be understood as the “fingerprint” of the image. We can effectively judge the degree of similarity of two images by the Hamming distance between phash codes. We can effectively filter the image dataset by using the phash algorithm and an appropriate threshold, which can substantially reduce the amount of calculation of matching process and improve computational efficiency. The second step of image retrieval is matching the color-orb descriptors to get the initial result, the binary codes match can use the XOR operation to seek Hamming distance which is very fast. The third step of image retrieval is using the geometric coding(GC) algorithm to verify the geometric relations between two images to get the final results. Experiments shows that, compared with traditional BOW and BOW+GC image retrieval systems, our method has a relatively good performance on search results, and has a great advantage on the retrieval time.