JatinArutla/Zero-Shot-Learning

Classifying unseen food dishes using Zero-shot learning achieving a Top-5 accuracy of 91%.

Classifying unseen food dishes using Zero-shot learning (ZSL) achieving a Top-5 accuracy of 91%.

Zero-shot learning is the process of solving a problem when there are no examples of that problem in the phase of training. Zero-shot learning methods are made to study various object classes, their features, and use the features learnt during image classification to help recognize unseen classes of data.

In this project, unseen food dishes are classified through Zero-shot learning. 9 training classes and 4 Zero-shot learning classes were considered in an attempt to classify the samples of ZSL classes using the Deep learning model trained on samples of training classes. The VGG16 model is used for image feature extraction of the zero-shot learning class samples and a new model is built for training on the samples of training classes. The word embeddings are gathered by using the pre-trained Word2Vecs by Google. The result of this is a Word2Vec for the 13 target categories that have been taken. After performing image feature extraction and normalization, the Top-5 classes are predicted using the Deep learning model. If the Top-5 predictions contain the actual label, then the model is said to have correctly classified the given image. The model achieved a Top-5 accuracy of 91% when tested on the ZSL classes.