Fusion of Handcrafted and Deep Convolution Networks Learned Features for Image Retrieval

The rapid increase in online services such as image sharing has led to the accumulation of unprecedented amounts of multimedia entities in repositories. Consequently, large amount of raw image data need to be analyzed for content and semantics. In this context, Content-Based Image Retrieval (CBIR) has become a crucial task in computer vision, focusing on retrieving images from large datasets based on the corresponding visual content rather than metadata. In this research, we propose a hybrid CBIR framework that integrate handcrafted and convolutional feature extraction and matching. The features extracted include Color Histogram, Oriented FAST and Rotated BRIEF (ORB), and deep features derived from a pre-trained ResNet-50 convolutional neural network. Handcrafted features capture information related to local, global, and computationally efficient keypoint descriptors, providing valuable insights into visual data, while deep features contribute discriminative properties for robust image representation. We fuse these complementary features to create a comprehensive image representation, enhancing retrieval accuracy and nominal retrieval time. Experimental analysis on benchmark datasets imply that the proposed approach satisfactorily outperforms traditional CBIR methods, achieving 96.6% high precision, effective recall and nominal retrieval time. Thus, the work highlights the efficacy of combining handcrafted and deep learning features for robust image retrieval applications.