I. Overview
Content-Based Image Retrieval (CBIR) - This is a long-established research area that requires improving the efficiency and accuracy of real-time retrieval. Artificial intelligence has made progress in CBIR, greatly facilitating the instance search process. In this survey, we review recent work on instance retrieval based on the development of deep learning algorithms and techniques, and the survey is organized by deep feature extraction, feature embedding and aggregation methods, and network fine-tuning strategies. Our survey considers a wide variety of state-of-the-art methods, through which we identify landmark work, reveal connections among various methods, and present commonly used benchmarks, evaluation results, common challenges, and Promising future directions are suggested.
Content-based image retrieval (CBIR) is the problem of searching for relevant images in an image library by analyzing the visual content (color, texture, shape, objects, etc.) given a query image. CBIR has been a long-standing research topic in the fields of computer vision and multimedia. With the exponential growth of image data, it is crucial to develop appropriate information systems that can efficiently manage such large image acquisitions, and image search is one of the most indispensable technologies. Therefore, the application potential of CBIR is almost endless, such as person/vehicle re-identification, landmark retrieval, remote sensing, online product search, etc.
In general, CBIR methods can be divided into two distinct tasks: category-level image retrieval (CIR) and instance-level image retrieval (IIR) .
The goal of CIR is to find arbitrary images that represent the same category as the query (e.g., dog, car). In contrast, in the IIR task, given a query image of a specific instance (e.g., the Eiffel Tower, my neighbor’s dog), the goal is to find images containing the same instance, which can be taken under different conditions, such as Different imaging distances, viewing angles, backgrounds, lighting and weather conditions (re-identifying examples of the same instance).