Dual-Phase Content-Based Image Retrieval System using both Handcrafted and Deep Learning Features

Abstract

Content-based image retrieval (CBIR) is a key technique for quickly getting and finding images from huge datasets by using image’s visual content. This study describes a dual-phase CBIR system that blends handcrafted features with features based on convolutional neural network deep learning model. The system comprises two phases and each phase involves feature extraction as well as feature matching. In the first phase, handcrafted-based features such as color statistics, Histogram of Orientated Gradients (HOG), Local Binary Pattern (LBP), and Discrete Cosine Transform (DCT) are extracted and then fused. In the second phase, deep learning-based features are extracted from the pre-trained ResNet50 model. Consequently, the extracted features are subsequently contrasted to the extracted features from the query image using Euclidean and cosine distance metrics in the first and second phases, respectively. Comprehensive experiments are carried out on the publicly available dataset of images known Corel-1K and precision rate of 95.16% and 96.21% are achieved for the top-20 and top-10, respectively. The findings of the experiment demonstrate that the proposed dual-phase strategy improves retrieval precision, beating both handcrafted-based and deep learning-based CBIR systems. Keywords:  CBIR, Deep learning, Handcrafted, ResNet50, Image similarity.