Improving corruption robustness of deep neural network in video classification

• Main Author: Gong, Zeyu
• Other Authors: Alex Chichung Kot
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/167077
• Institution: Nanyang Technological University

Recently, video social media platforms grow fast and wide, video analysis become a necessary task. Deep neural networks have previously demonstrated exceptional performance on a variety of computer vision tasks, including image classification and action recognition. However, some studies prove that when the training video clips and test video clips are marginally different, the accuracy of the deep neural network can observably decrease. And corrupted test videos can lead to even worse test accuracy. The corruption robustness of a video classification model refers to its capability to maintain accurate predictions when presented with corrupted input data. The corruption can occur due to various factors such as weather, sensor noise and transmission errors. Therefore, the robustness of the video classification model is still a problem and there is a large progress to be made. In this project, we use two popular video datasets which have various classes of videos and different human actions as the train datasets. We train a video classification model with the clean dataset and evaluate its robustness with test datasets. We apply different corruptions to the original fine videos to serve as test datasets. To test the robustness, we test the model with corrupted videos and evaluate the accuracy of its classification result. Based on the test and evaluation result, we aim to train a deep neural network that has good corruption robustness compared with previous models or make improvements to the robustness of existing video classification models.