A Deep Convolutional Neural Network for Interrelationship Identification between Humans from Images

Abstract

The paper proposes a deep convolutional neural network for visual categorization of different interrelationships between humans from digital images. To achieve this goal, we first generated a dataset of interrelationships containing two interrelationship classes i.e. handshaking and hugging. Our proposed network having around 2 lakh neurons is trained with 8 lakh parameters. The network contains a total of seven layers i.e. two convolution layers each followed by max pooling layers and fully connected layers. Output layer contains a sigmoid function providing binary outputs i.e. 0 for one class and 1 for another class. To maintain the nonlinearity of images, Rectified Linear Units (ReLUs) have been used in each convolution and fully connected layers. The model generates an average accuracy of approximately 81%. Data augmentation technique has also been applied to reduce over-fitting.