ABSTRACTThis paper presents a technique for content-based image retrieval (CBIR) by selecting the regions on the basis of their contribution to image contents. We have analyzed the problems associated with matching regions among the pair of images over the large set of overlapping regions. It is being studied that matching images by using regions having unstructured association can be a serious problem. In this research, we propose a linear formulation technique, which involves simultaneous matching, so that the matched area can have color-similarity histogram, shape and having little overlapping region. It is also analyzed that the selected region can have a small number and overall concavity is low, and tried to cover both the images. It has been studied that CBIR has attracted many researchers and most of the previous CBIR systems have shown the searching procedure of the digital image on the basic features such as texture, color, size, and shape of a certain query image in a large database. According to this research, we are going to present a region-based image repossession system that is going to exhibit a model that would help specify multiple regions of interest inside the query image. In this research, we have presented a novel visual feature that might contain color size of the region query and its moments, however, to combine color and region-size information of the watershed region. Moreover, a technique has been modeled for region filtering that might depend upon the color size of the given query image and that would stimulate the process of screening out the most nonrelated region and images for pre-processing of the recovery of image. Therefore, the technique presented would help shorten the consequence of image background on image-matching decision; however, an object's color would receive much more focus. Apart from that, amendment to region-based similarity measurement has also been presented. It has been proved with the help of simulation results that the given descriptor with the similarity measure amendments outperforms the existing descriptor that would be considered in content-based image-retrieval applications. Moreover, the given approach has performed better than the previous approaches. Our method would be based upon a simple and reliable metrics, which is being used to calculate similarities. We have performed numerous simulations and verified that the given approach has outperformed the current techniques in localization, and is going to have vigorous object discovery in the existing mixed-class dataset.