Large-Scale Image Action Recognition Depends on Deep Learning

Wei Wu, Jiale Yu

Abstract: In the field of computer vision, research on action recognition in still images has begun to attract people's attention. However, the single action image contains limited information and is difficult to identify. Therefore, the action recognition of still images still faces the enormous challenges. In order to solve this problem, this paper uses the 34-layers residual convolutional neural network to construct recognition model. By extracting the features of the last fully connected layer from this CNN model, we forward the features to the classical models such as Decision Tree, Random Forest, Bayes, and SVM. Experimental results on our images datasets show that our method achieves better performance than a single CNN and other traditional classifiers, especially when combined our residual network with SVM, the recognition accuracy is over 80%. It confirms the effectiveness of our method.