A Dual-Task Deep Neural Network for Scene and Action Recognition Based on 3D SENet and 3D SEResNet

Abstract

Aiming at the problem that scene information will become noise and cause interference in the feature extraction stage of action recognition, a dual-task deep neural network model for scene and action recognition is proposed. The model first uses a convolutional layer and max pooling layer as shared layers to extract low-dimensional features, then uses 3D SEResNet for action recognition and 3D SENet for scene recognition, and finally outputs their respective results. In addition, to solve the problem that the existing public dataset is not associated with the scene, a scene and action dataset (SAAD) for recognition is built by ourselves. Experimental results show that our method performs better than other methods on SAAD dataset.