A Comparative Study of Pre-Trained CNN Models with Transfer Learning for Content-Based Image Retrieval

Content-Based Image Retrieval (CBIR) involves searching for images that are visually similar within large image databases. Convolutional Neural Networks (CNNs) are important in CBIR tasks and classifications. However, handling large-scale datasets with a high number of categories continues to be a significant challenge. This study investigated the effectiveness of four widely used pretrained models, EfficientNet-B0, DenseNet-201, VGG-16, and AlexNet, for CBIR on the Corel-10K dataset. The four models are fine-tuned using transfer learning and data augmentation techniques and evaluated using mean Average Precision at rank K (P@K) and across all queries (mAP@K) to enhance their retrieval performance. Experimental results prove that the DenseNet-201 architecture achieves the best retrieval performance among the models, with a P@10 of 92.1% and a mAP@10 of 94.7%. The VGG-16 model also performs well, while EfficientNet-B0, despite showing lower performance compared to others, maintains computational efficiency. DenseNet-201 achieved superior precision on the Corel-10K dataset compared to other methods, demonstrating its effectiveness in large-scale image retrieval. This study helps identify effective pre-trained models for CBIR and serves as a basis for future advances in retrieval systems, including better handling of image orientation changes.