Action Recognition in Still Image Using Deep Neural Network

Hoo, Min Jia (2020) Action Recognition in Still Image Using Deep Neural Network. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Hoo Min Jia.pdf Restricted to Registered users only Download (3MB)

Abstract

Action recognition in still images is a challenging topic as images do not consist of motion information like in videos. The existing works of still image-based action recognition are reviewed. The basic flows of action recognition by using deep learning are image preprocessing, feature extraction and classification. The proposed method of action recognition in still images uses the different neural networks including VGG-16, ResNet-50 and Inception-V3 as feature extraction. The proposed method are tested by using the dataset, Stanford-40. Training accuracy and validation accuracy for VGG-16 are 99.63% and 82.8% respectively and it achieved the test accuracy of 56.76%. 99.44% and 92.07% accuracy for the training and validation but the test accuracy is the lowest which is 47.84%. Inception-V3 achieved As for ResNet-50, the training accuracy and validation accuracy are 98.41% and 88.41% respectively and it achieved the test accuracy of 59.09% which is the highest among these networks.