Imbalanced Dataset Visual Recognition by Inductive Transfer Learning

Abstract

Computerized picture is a blend of picture components called pixels, each with a discrete and finite numeric value for its intensity level. Image classification is the process of extracting this pixel features in a digital image and based on this features the image is assigned into a class category. Transfer learning has been quite successful in processing image classification task with less number of training images. Motivated by the immense application of transfer learning methodology, we propose a novel architecture for classifying fine-grained and low resolution images with inter-class and intra-class similarity. In the proposed model we are examining the accuracy level of knowledge transfer between source and target domains having heterogeneous and homogeneous feature spaces and label spaces. The datasets we are using are Oxford 102 flower dataset and Caltech 101 dataset. The Inductive transfer learning methodology is being adapted. As the available target dataset per class is less, we are emphasizing few-shot learning. Supervised learning method is used for feature construction. ResNet50 deep neural networks are used as fixed feature extractors. The accuracy exhibited by proposed architecture is compared with the state-of-the-art image classification models.