Content-Based Image Retrieval Established on Deep and Handcrafted Features

Abstract

The role of image features is crucial in any system. There is a need to enhance Content-Based Image Retrieval (CBIR) systems, but challenges exist in accurately classifying, retrieving, and browsing or mining images. These challenges can be effectively addressed through the extraction of meaningful visual features. Various handcrafted and Deep Learning (DL) techniques have been developed for this purpose, but there has been limited exploration of combining the two approaches. The proposed technique is based on joint use of handcrafted features, i.e., Histogram of Oriented Gradients (HOG), Speed-Up Robust Features (SURF), Bag of Features (BoF), and Local Binary Pattern (LBP), and deep features extracted through AlexNet plus Spatial Pyramid Pooling (SPP) model. The Support Vector Machine (SVM) classifier was used for the classification. The algorithm was evaluated using the Caltech-256 RGB image dataset, achieving an average accuracy of 86.8%. The outcomes demonstrated the benefits of combining handcrafted and DL features, leading to improved accuracy in specific CBIR scenarios.