Enhanced Zero shot Learning using Deep Neural Network ResNet50

Computer vision has advanced with lot of development in visual recognition systems, which poses restrictions to expand for huge numbers of image classes. This restriction is due to huge image classes with unlabeled images which cannot be used in training the machine learning algorithm. As Traditional machine learning method of classification are based on the classification of categories which are available at the time of training. Technique of Zero-shot learning (ZSL) recognizes categories of test sets which is not appearing while training the model. The enhanced ZSL technique proposed is based on deep visual semantic embedding method. In this method Visual and semantic features are used for the classification of unknown categories. The extraction of visual features is accomplished with convolutional neural network (CNN) and ResNet 50. FastText is used to convert labels of classes into word embedding vectors. Visual and word embedding features are mapped. The model is predicting the Top 5 labels for an unknown image category (zero-shot class). The experiments are performed on standard datasets SUN and AWA2. The proposed technique of enhanced ZSL with ResNet 50 gives better accuracy and reduced model loss then CNN model.