The Research Of Image Retrieval Method

With the technique of network and storage developing so rapidly, the quantity of image is increasing quickly. There are more and more images spreading on the Internet resulting in the problem of how to quickly finding the desired images. Users want to quickly,easily and efficiently find their target images in the vast amount of source. Content based image retrieval (CBIR) was proposed to resolve the problem of impossibility of manual annotation on tedious image data and the manual annotation also being subject to human beings. At the same time, CBIR technique uses the computer to automatically extract the features of images for retrieval process. This is very important point now because of the vast amount data source.Content based image retrieval first uses the computer to automatically extract the low-level features of images to describe the semantic meanings of the images, then defines the similarity metric on the feature space and computes the similarity between the images in the database and the query image. CBIR was indeed successful at the first years. Many commercial and research systems have being developed such as QBIC and Informedia, etc. But because of the internal disadvantage of CBIR, "Semantic Gab" --the gab between the low-level features and high level features. The low-level features extracted from the images are easily subject to the variance of luminance,colour and post-process and so on, while the high-level semantic meanings of images are not subject to these changes.To overcome this "Semantic Gap" problem, researchers are beginning to find the other techniques to bridge this gap. The region based image retrieval,relevance feedback and high-level semantic meaning modelling techniques have being proposed. The method of region based image retrieval and high-level semantic meaning modelling are not through different with CBIR technique because they still use the low-level features of images for the retrieval process. While at the same time, they use some ad-hoc machine learning,graph model and other methods to get a better classifier through training on the source data. The relevance feedback technique different from the above two, use a new technique to improve the image retrieval performance. This technique considers users as an important part of image retrieval. The system can use the information that users feeding back in the retrieval process to improve the retrieval performance. What other information can we use beyond the pixel of the image? That is camera metadata—Exif, which can be used to improve the understanding of high-level semantic meanings of images. The Exif information includes all parameters values of camera such as: the Flash turn on/off,the Exposure time,the Focal length and the F-number, etc. These information values are valuable for the semantic meanings of images. The on going research has already proved this point. Different semantic meanings have different relationship with the camera metadata values, such as: the Flash being more frequently used in the indoor concept while not outdoor concept; the value of Focal length to the scene concept and object concept being different too. The problem here is how we find this relationship.This paper summarizes the above methods and proposed a region based image retrieval method incorporated with camera metadata. This method first uses the image segmentation method to segment the image into region blocks, and then extracts the low-level features of MPEG-7 descriptor based on the above region blocks. According the defined similarity, we treat the retrieval results using above low-level features as raw results. Then we fusion the camera metadata information automatically extracted from the images into the difference equation of two images. We use the F-number and Focal length of camera metadata as the post-process information of image retrieval to improve the retrieval results. The experiment results show that our proposed method gets the better performance results on the car,garden, grass,high-tower, sculpture and red-building six concepts compared with the traditional content based image retrieval method.