Points Of Interest Based Image Retrieval

In content-based image retrieval, global features related to color or texture and shape are commonly used to describe the image content. The problem with this approach is that these global features cannot capture all parts of the image having different characteristics and can't contain the spatial features. On the other hand, user could be interested to retrieval only similar parts or object of images. In this context, global image descriptors are unusable. Therefore, local computation of image information is necessary to improve retrieval result.In order to make up for the limitation of global features, each region of the image must be analyzed independently. More precisely, the concept consists first in localizing relevant regions in the image and second in characterizing relevant primitives obtained with local descriptors. Region of interest is one of the solutions. A region of interest is a portion of image that you are very interested in and holds the main information of the image. But this method depends on the good image-segment algorithm. The perfect image-segment algorithm isn't found now. Schmid and Mohr introduced the notion of interest points which are used in computer vision in image retrieval. First, they localized relevant regions by detecting interest points in the image. Then a small region around the interest point is located as an image patch. Low-level features are extracted to describe each image patch. Finally, the similarity measurement can be done by using interest points matching. Interest points represent the image details flexibly and are robust to geometric transformations of the image. Further more this method avoids image-segment.On the basis of Schmid and Mohr's researches we improve the interest points based image retrieval. The main contributions of this paper are:1. We extract color moments to describe each image patch and transform RGB color space into HSV color space. Partial illuminance invariant can be achieved by setting all the weights for the moments of the hue channel to a higher value than the other weights.2.The geometry hashing technic is used to enhance the geometry constraint of interest points in order to improve the matching result. The correct matching of interest points is very important to the retrieval result. The geometry hashing technic not only insures the correct matching of interest points but also makes the retrieval result more roust to rotation, translation and scaling.3. We improve the voting algorithm during the similarity measurement. The difficulty of selecting threshold of voting is decreased and the retrieval result is improved by using weighted voting technic.4.Experiments show that this method is very effective by our experiment system of interest points based image retrieval.