Synergistic fusion of handcrafted descriptors and deep convolutional neural network features for improved content-based image retrieval

Content-Based Image Retrieval (CBIR) is vital to the orderly organization and effective retrieval of visual information across a variety of areas such as healthcare, surveillance, e-commerce and remote sensing. This paper proposes a synergistic CBIR framework that fuses handcrafted descriptors such as Local Binary Patterns (LBP, 256D), Histogram of Oriented Gradients (HOG, 1764D), Scale-Invariant Region-based Features (SIRF, 512D), and Color Moments (9D) with deep CNN features extracted from a fine-tuned VGG16 (1024D) and a pretrained ResNet50 (2048D). Unlike conventional hybrid fusion strategies, the proposed approach achieves a synergistic effect, wherein the joint representation captures both fine-grained local structures and high-level semantic abstractions more effectively than individual components. The full hybrid vector (5613D) is reduced through supervised LightGBM-based feature selection (K = 2048) followed by PCA (d = 128, 97.3% variance retained), and indexed using FAISS-HNSW for efficient approximate nearest-neighbour retrieval with cosine re-ranking. Experiments on a combined dataset of 19,632 images across ten semantically coherent meta-classes demonstrate: Precision@20 = 0.9820, Recall@20 = 0.8290, F-measure = 0.8700, and overall Accuracy = 97.89%–outperforming handcrafted-only (Precision@20 = 0.9264), deep VGG16-only (0.8745), deep ResNet50-only (0.9163), and recent state-of-the-art hybrid methods. Comprehensive computational efficiency analysis shows an average query time of 1.2 ms and throughput of 892 images/second on an NVIDIA DGX-H100.