A novel content-based image retrieval system using custom CNN and data augmentation for improved rotation robustness

Content-Based Image Retrieval (CBIR) is essential for retrieving images through visual content comparison, addressing the limitations of traditional keyword-based searches. CBIR systems face challenges in addressing the semantic gap between low-level features and high-level human concepts and handling rotation-based retrieval issues. This paper aims to create a novel CBIR system that combines a custom Convolutional Neural Network (CNN) architecture with data augmentation techniques for a rotation-invariant image retrieval system. The primary focus of this work is to obtain high precision and build a rotation-invariant system. The custom CNN architecture is designed to extract both detailed low-level features and high-level semantic information from images. To enhance robustness against rotations, the system integrates data augmentation techniques during the CNN training phase. This leads to the accurate retrieval of relevant images even when rotated. The datasets used in this research were collected from Kaggle online sources. The proposed system achieves an overall average precision of 98.5% on the Corel-1K dataset and 94.3% on the natural image dataset, showing significant improvements over existing methods evaluated on the same datasets. This work advances CBIR by introducing a custom CNN that enhances retrieval precision and ensures rotation invariance, setting a new standard for future image retrieval systems.