Enhanced Low-Level-Feature- and Color-Aware Content-Based Image Retrieval Using Deep Learning

With the rapidly growing number of images being captured by digital devices, e.g., cameras and mobile phones, the demand for finding similar images is increasing.However, traditional techniques that rely on investigating low-level features, such as edges, corners, and blobs, or color histograms, have become obsolete, as such information may be similar in unrelated images.Moreover, methods that rely on matching high-level features, e.g., objects in the image, may also produce false matches, as unrelated images may also contain similar objects in them.This gap in matching images is known as the semantic gap, which is caused by the reliance on low-level or high-level features in the images to match them.In this paper, a novel similarity measure is proposed to enhance the performance of contentbased image retrieval.Initially, three similarity measures (high-level features, low-level features, and colors) are calculated per pair of images, which are then multiplied by three factors and used to adjust the weight of each of these similarities using the PSO algorithm and fused to a single similarity measure.Next, a database is generated based on the proposed similarity measure, in which the similarity measure of each pair of images is used as the training value for a neural network that takes the pair of images as input.Finally, we used ANNs to approach the complex computations of the proposed method, which allows the proposed method to be used in real-life applications.This model implementation is based on the Yolo V8 neural network.The proposed method is compared with recent stateof-the-art methods using the MIRFLICKR-25K, NUS-WIDE, MS-COCO, Pascal VOC2007, and Pascal VOC2012 datasets in terms of three well-known metrics (NDCG, ACG, and wMAP).As a result, the proposed method has been able to outperform all existing methods in terms of NDCG, with the highest of 0.98.These results illustrate the importance of the inclusion of the three similarity factors, the high-and low-level features as well as the color histogram, in retrieving more-relevant results to the query image, in addition to the ability of neural networks to approximate complex computations to reduce execution time while maintaining high-quality results.