ABSTRACT In today's rapidly evolving digital landscape, the demand for multimedia applications is surging, driven by significant advancements in computer and storage technologies that enable efficient compression and storage of visual data in large-scale databases. However, challenges such as inaccuracy, inefficiency, and suboptimal precision and recall in image retrieval systems necessitate the development of faster and more reliable techniques for searching and retrieving images. Traditional retrieval systems often rely on RGB colour spaces, which may inadequately represent critical image information. In response, we propose a content-based image retrieval (CBIR) system that integrates advanced techniques such as quadtree segmentation alongside modern lightweight deep learning models, specifically MobileNet and EfficientNet, to enhance precision and recall. Our comparative experiments reveal that these deep learning models significantly outperform traditional methods, including SVM classifiers combined with feature extraction techniques such as Histogram of Oriented Gradients (HOG), Scale-Invariant Feature Transform (SIFT), and Speeded-Up Robust Features (SURF). Notably, MobileNet and EfficientNet achieved F1-scores of 0.87 and 0.89, respectively, with enhanced processing efficiencies that resulted in feature extraction times reduced to 20 ms and classification times down to 8 ms. This translates to rapid image retrieval times as low as 35 ms, highlighting the superior performance of modern deep learning models in enhancing both retrieval accuracy and efficiency for large-scale image databases, making them ideal for real-time applications.
Read full abstract