Publication

Abstract : A blend of unlabeled data (unsupervised) in large volume, along with a small number of labeled data (supervised), makes a semi-supervised dataset. Content-Based Image Retrieval (CBIR) systems are used to label/model such data. Based on the features extracted, the unsupervised query images get labeled upon the similarity measures with the reference labeled images (one image per class). Due to the insufficient availability of trained data, modeling a CNN to act as a feature extractor will be inefficient. Hence, pre-trained models, trained upon public databases (ImageNet) and have never been fine-tuned on the target dataset, will be used as a feature extractor, which lowers the confidence over the labeling task. The proposed method will put together a Generative Adversarial Network (GAN), with the chosen pre-trained model embedded inside the GAN as a discriminator, that will fine-tune the pre-trained model's weights in accordance with the unlabeled target data. These fine-tuned pre-trained models can then be used to extract features and will help to build a robust CBIR system. Abbreviations - CBIR, Content Based Image Retrieval; GAN, Generative Adversarial Network; CNN, Convolutional Neural Network.