Content based Image Retrieval using Fine-tuned Deep Features with Transfer Learning

This paper introduces a deep learning approach to efficiently retrieve images using a robust deep features extracted from VGG-19 architecture. Our work involves fine-tuning this pre-trained network to adapt it to our dataset by replacing the final layers, we can then train the network so that it learns feature representation for the content based image retrieval task. After applying transfer learning, then, the new network is trained on our dataset after performing augmentation to the set of training images to increase the number of images in order to improve the training accuracy. Augmentation techniques involve using shifting, shearing, and flipping to the original input images. Finally, features are extracted from the ‘fc7’ layer that has 4096 bins for each input image. Euclidean distance is applied to calculate the closest distance between the query image and the features database. Experiments are conducted on a standard dataset called Corel-1k with 1000 images and 10 different categories. Results show that our approach generates high precision accuracy that outperforms traditional image retrieval methods and is in level with deep learning based methods.