This study presents CTD-Net (Color, Texture, and Deep Learning- Network), a novel Content-Based Image Retrieval (CBIR) system that integrates handcrafted and deep learning features to achieve enhanced retrieval performance. Traditional CBIR systems often rely solely on either handcrafted features or deep learning models, leading to limitations in capturing the complex visual and semantic information of images. CTD-Net addresses this issue by incorporating color features derived from Color Histogram and Color Moments, texture features extracted via Local Binary Patterns and Wavelet Transform, and deep features from the EfficientNet-B7 model. The fusion of these features enables CTD-Net to bridge the gap between low-level visual qualities and high-level semantic information, resulting in improved precision. Experimental results on the Corel-1 K, 10 K and Caltech-101 datasets demonstrate that CTD-Net achieves precision rates of 98.85%, 92.40%, and 88.94% respectively, significantly outperforming existing methods. The proposed approach not only highlights the effectiveness of hybrid feature fusion but also underscores its potential to advance CBIR systems. The source code of this research is available at https://github.com/Surbhityagi/CTD-Net-for-CBIR.git.
Tyagi et al. (Tue,) studied this question.