Image Feature Extraction for Similarity Searching Using Transfer Learning with ResNet

Abstract

This study evaluates the use of feature extraction with a pre-trained ResNet50 model for similarity search tasks. We employed transfer learning from both initial and intermediate layers of ResNet50 and applied a robust preprocessing approach, including resizing and Gaussian blur, to optimize feature extraction. The method achieved over 90% accuracy in nearest neighbor (NN) and k-Nearest Neighbors (k=3 and k=5) searches for both logos and paintings datasets. Notably, our approach does not require fine-tuning, which is advantageous when only one instance of each element is available. Overall, the method is effective and practical for similarity search applications.