DFF-CBIR: Deep Feature Fusion Framework for Enhanced Content-Based Image Retrieval

Content-Based Image Retrieval (CBIR) has become an essential component in a wide range of vision-driven applications, where the objective is to retrieve semantically similar images from large-scale repositories based on visual content. In this work, we propose an enhanced hybrid feature extraction framework that combines ConvNeXt and EfficientNet for richer representation learning. Unlike conventional single-model approaches, our method concatenates the layer embeddings from the two networks, enabling a more comprehensive capture of local texture descriptors and global semantic abstractions. To rigorously assess similarity, we incorporate a broad spectrum of distance measures, including Cosine, Euclidean, Manhattan, Chebyshev, Minkowski, and Canberra. Experiments are conducted on three diverse and widely used benchmark datasets: Corel-1K, Corel-10K, and Caltech-101. The proposed dual-branch feature fusion consistently surpasses the individual backbone models and baseline methods across multiple evaluation metrics. Specifically, the framework achieves a precision of 99.17% on Corel-1K, 90.74% on Corel-10K, and 92.24% on Caltech-101, along with strong recall, F1-score, and accuracy performance across all distance measures. These results validate the robustness of the hybrid representation in capturing inter-class variability and intra-class similarity. This study contributes a scalable, architecture-agnostic paradigm for deep feature fusion in CBIR and highlights the potential of enriched hybrid embeddings for improving visual search in large-scale image databases.