Opt-IncpResN-V2: A Hyper-Optimized Framework for Image Quality Enhancement and Feature-Rich Content-Based Image Retrieval

Content-Based Image Retrieval (CBIR) has been rapidly developed and implemented in different sectors connected to computer vision and artificial intelligence due to the growth of the Internet and the prevalence of digital devices. An input picture can efficiently and effectively retrieve related images from a large-scale database. Over the last ten years, much work has been done to develop new CBIR theories and models, and numerous efficient CBIR algorithms have been developed. Finding identical pictures in an extensive collection that match the query picture is made more accessible through content-based image retrieval. Ranking the pictures for retrieval usually involves considering the representative feature similarities between the dataset and query pictures. This study introduces Deep Learning (DL)-based Opt-IncpResN-V2, a hyper-optimized framework designed to enhance image quality and enable feature-rich CBIR. With the proliferation of digital images and the expansion of image databases, efficient image retrieval methods are essential. Unlike traditional text-based image retrieval (TBIR), which relies on manual keyword annotation, CBIR retrieves images based on visual features such as colour, texture, and shape. Our framework incorporates an image quality improvement process during preprocessing, including noise reduction, edge detection, normalisation, and contrast enhancement. Features are extracted using the Inception-ResNet-V2 technique, and classification accuracy is improved using the Artificial Flora Optimisation Algorithm (AFOA). Images are then classified using DenseNet-264, and hyperparameters are optimised using the Improved Hippopotamus Optimisation Algorithm (IHOA). The proposed approach is evaluated on four benchmark datasets: Kvasir, CIFAR-10, CIFAR-100, and Corel-1k. Experimental results demonstrate superior performance compared to state-of-the-art techniques, achieving accuracy, precision, recall, and F1-scores exceeding 99% across all datasets.