This study presents a comparative analysis of state-of-the-art deep learning models–EfficientNetB0, MobileNetV2, and ResNet101–for image classification and content-based retrieval in scientific publications. A dataset of 4,303 images from 11 categories was curated from the Can Tho University Journal of Science and enhanced through tailored data augmentation strategies. The models were fine-tuned using transfer learning with hyperparameters optimised via Grid Search. Features were extracted using GlobalAveragePooling2D, and cosine similarity was combined with the FAISS library for efficient similarity search. Experimental results demonstrate a clear performance-efficiency trade-off: ResNet101 achieved the highest classification accuracy, while EfficientNetB0 and MobileNetV2 offered significant advantages in inference speed. A user-friendly web interface was developed to support practical image retrieval applications. These findings highlight the potential of deep learning in enhancing the management and integrity of scientific image resources.