I.J. Image, Graphics and Signal Processing, 2023, 2, 1-12 Published Online on April 8, 2023 by MECS Press (http://www.mecs-press.org/) DOI: 10.5815/ijigsp.2023.02.01 This work is open access and licensed under the Creative Commons CC BY License. Volume 15 (2023), Issue 2 A Novel Technique for Image Retrieval based on Concatenated Features Extracted from Big Dataset Pre-Trained CNNs Chandra Mohan Bhuma Bapatla Engineering College, Department of ECE, Bapatla, Andhra Pradesh, India E-mail: chandrabhuma@gmail.com ORCID iD: https://orcid.org/0000-0002-7566-4739 Ramanjaneyulu Kongara* PVP Siddhartha Institute of Technology, Department of ECE, Vijayawada, Andhra Pradesh, India E-mail: kongara.raman@gmail.com ORCID iD: https://orcid.org/0000-0003-0711-0547 *Corresponding Author Received: 08 April, 2022; Revised: 14 June, 2022; Accepted: 10 February, 2023; Published: 08 April, 2023 Abstract: Accessing semantically relevant data from a database is not only essential in commercial applications but also in medical imaging diagnosis. Representation of the query image by its features and subsequently the dataset are the key factors in Content Based Image Retrieval (CBIR). Texture, shape and color are commonly used features for this purpose. Features extracted from the pre-trained Convolutional Neural Networks (CNNs) are used to improve the performance of CBIR methods. In this work, we explore a recent state of the art big dataset pre-trained CNNs which are known as Big Transfer Networks. Features extracted from Big Transfer Network have higher discriminative power compared to features of many other pre-trained CNNs. The idea behind the proposed work is to demonstrate the effectiveness of using features of big transfer networks for image retrieval. Further, features extracted from big transfer networks are concatenated to improve the performance of the proposed method. Feature diversity supplemented with network diversity should ensure good discriminative power for image retrieval. This idea is supported by performing simulations on four datasets with varying sizes in terms of number of images and classes. As feature size increases with the concatenation, we applied a dimensionality reduction algorithm i.e., Principal Component Analysis. Several distance metrics are explored in this work. By properly choosing the pre-trained CNNs and distance metric, it is possible to achieve higher mean average precisions. ImageNet-21K pre-trained CNN and Instagram pre-trained CNN are chosen in this work. Further, a pre-trained network trained on Imagenet-21K dataset is superior compared to the networks trained on ImageNet-1K dataset as there are more classes and presence of wide variety of images. This is demonstrated by applying our algorithm on four datasets i.e., COREL-100, CALTECH-101, FLOWER-17 and COIL- 100. Simulations are presented for various precisions (scopes), and distance metrics. Results are compared with the existing algorithms and superiority of the proposed method in terms of mean Average Precision is shown. Index Terms: Content Based Image Retrieval, Big Transfer Network, Pre-trained Convolutional Neural Network, ImageNet-21K, COREL-100, CALTECH-101, FLOWER-17, COIL-100 datasets. 1. Introduction Based on the content of the query image, extracting relevant images from the large databases is known as Content Based Image Retrieval (CBIR) [1]. Image search engines utilize this for displaying images when the user poses a query. Finger print recognition, iris recognition, crime detection and abnormality assessment from the medical images are some applications of CBIR. Two steps in general CBIR systems are feature extraction and similarity measure. Feature extraction refers to representation of the images in a compact manner without any loss of image content. Several features i.e., texture, color, and shape [2] are used for this purpose. In recent times, Convolutional Neural Networks (CNNs) are gaining popularity due to their powerful representation of the images with minimal preprocessing. Several architectures have been attempted using CNNs and excellent performance metrics have been reported on various