Action Recognition in Still Images using Residual Neural Network Features

Abstract

Action recognition in still images is central to various problems like object recognition, image retrieval, scene recognition, pose estimation and understanding of actions in videos. Human actions and behaviour in images are recognized in this problem. Images have no spatiotemporal features. So the process of action classification in videos is not applied to images. The existing works of still image action classification are reviewed. An image action recognition model is proposed in which the feature extraction is done using a deep neural network such as residual neural network and Support Vector Machine(SVM) acts as the classifier. The proposed model is evaluated with standard benchmark datasets like Pascal VOC action dataset, and Stanford actions 40 datasets using the evaluation measures such as precision, recall,f1-score, and Kappa measure. The model has high performance in Pascal VOC dataset. The model is outperformed in fifteen classes and achieved the best performance compared to other works.