Exploring Pretrained Models and Transfer Learning Techniques for Image Retrieval

In recent years, the field of content-based image retrieval (CBIR) has witnessed significant advancements and garnered immense attention. With the increase of digital imagery across various domains, the need for effective image search and retrieval techniques has become increasingly important. Pre-trained deep neural networks have arisen as effective tools for image analysis and feature extraction due to their ability to learn complex visual representations from vast amounts of data. By harnessing the power of pre-trained deep neural networks, such as AlexNet, ResNet, VGG16, and GoogLeNet, researchers seek to enhance CBIR systems. This paper explores the significance of utilizing pre-trained networks for content-based image retrieval (CBIR). It includes a literature review discussing similar CNN architectures & their applications as well as identifying research gaps. The paper conducts a comparative study of the four pre-trained networks: AlexNet, ResNet, VGG16, and GoogLeNet, evaluating their performance and suitability for CBIR. It also demonstrates transfer learning techniques to fine tune AlexNet and VGG16 models using MNIST and Fashion MNIST dataset. Evaluation using LRAP as a metric revealed good performance for both models, with AlexNet consistently outperforming VGG16, albeit with longer execution times.