An effective bi-layer content-based image retrieval technique

Abstract

Every second, millions of users across the world share and download an enormous quantity of multimedia content produced by different image capture devices. The substantial amount of computing is incurred to provide visually similar results to the user’s query. Content-based image retrieval (CBIR) is an automated mechanism for retrieving similar/closer images from an image collection depending on the extraction of the visual content from the images themselves. Research in this area generally involves two directions. The first direction focuses on the effectiveness of the description of the visual content of images, namely features, by a technique that leads to the discernment of similar and dissimilar images, and eventually the retrieval of the closer images to the query image. Meanwhile, effectiveness refers to improving the precision rate that leads to discriminate similar and dissimilar images accurately. The second direction concerns the retrieval efficiency in terms of time consumption. This paper mainly focuses on effectiveness rather than efficiency. Generally, there are two kinds of visual features: the global features and the local features. This paper takes advantage from both global and local features, and hence, a hybrid CBIR technique is developed in which it contains two layers of search (so-called bi-layer). The first layer aims to eliminate/exclude the dissimilar images, as much as possible, that leads to decrease the range of the search, by comparing all images in the dataset to the query image depending on the well-known local feature descriptor called Speed-Up-Robust-Features, which is a type of descriptor used in a bag of features technique for image clustering purposes. The second layer aims to compare the query image to the images attained/remained in the first layer on the bases of extracting global features like shape, texture, and color to retrieve the certain number of the similar/closer images to the query image. Additionally, this proposed approach also introduces the idea of producing the cases to make the system more dynamic in order to increase the precision rate. The performance of the proposed CBIR approach has been assessed utilizing the Corel-lK dataset. The experimental findings demonstrated the influence of exploring the concept of bi-layer in enhancing precision rate in contrast to state-of-the-art approaches, that reached 86.06% and 80.72% for top-10 and top-20, respectively.