EFFICIENT IMAGE DESCRIPTOR GENERATION USING CNN ARCHITECTURES FOR ENHANCED IMAGE RETRIEVAL

Abstract

Machine learning algorithms are widely employed in image classification tasks to extract and represent discriminative features from images. In this study, we present an efficient approach for generating image descriptors using Convolutional Neural Network (CNN) architectures, including GoogleNet, Inception V3, and DenseNet-201. These networks are leveraged to capture both texture and object-level features, which are further encoded through three color channels to enhance image retrieval performance while maintaining an optimal response time. When images are processed through the hierarchical layers of the CNNs, distinctive feature representations (signatures) are produced. These signatures are subsequently used to construct a new matrix that effectively encodes spatial relationships, color attributes, and latent patterns, thereby providing a more comprehensive representation of image content. The proposed CNN-based method was evaluated on four benchmark datasets: Corel-1K, CIFAR-10, 17-Flowers, and ZuBuD. Among the tested architectures, DenseNet-201 achieved the best performance on the CIFAR-10 dataset, which contains images of diverse categories and varying sizes, demonstrating superior accuracy compared to GoogleNet and Inception V3.