Advanced multilevel feature fusion framework for enhanced image retrieval using convolutional neural network and benchmark datasets

Abstract Despite deep learning helping image analysis many more data representation challenges remain along with performance consistency across different image types. Using convolutional neural networks (CNNs) architectures, a novel method is presented in this study that merges multilevel CNN features from top-performing networks like AlexNet, DenseNet, GoogLeNet, InceptionNet, and ResNet-101. The presented method starts with preprocessed inputs that move through CNN architecture layers for feature extraction while fusing output from various models to classify results across different testing datasets. The presented framework undergoes testing on seven datasets such as Tropical-Fruits, 101-ObjectCategories, CIFAR-10, ALOT, Corel-10k, 17-Flowers, and Zubud to confirm its usage across various scenarios. It uses top 10 to top 50 retrieval evaluations to demonstrate quick and precise image retrieval performance. The presented method consistently achieves superior results throughout multiple tests with high-quality image retrieval accuracy rates alongside effective classification and flexible use across various real-world situations.