Master of Science (MS)
With the growing demand for gender-related data on diverse applications, including security systems for ascertaining an individual’s identity for border crossing, as well as marketing purposes of digging the potential customer and tailoring special discounts for them, gender classification has become an essential task within the field of computer vision and deep learning. There has been extensive research conducted on classifying human gender using facial expression, exterior appearance (e.g., hair, clothes), or gait movement. However, within the scope of our research, none have specifically focused gender classification on two-dimensional body joints. Knowing this, we believe that a new prediction pipeline is required to improve the accuracy of gender classification on purely joint images. In this paper, we propose novel yet simple methods for gender recognition. We conducted our experiments on the BBC Pose and Short BBC pose datasets. We preprocess the raw images by filtering out the frame with missing human figures, removing background noise by cropping the images and labeling the joints via the C5 (model applied transfer learning on the RestNet-152) pre- trained model. We implemented both machine learning (SVM) and deep learning (Convolution Neural Network) methods to classify the images into binary genders. The result of the deep learning method outperformed the classic machine learning method with an accuracy of 66.5%.
Sung, Cheng-En, "GENDER CLASSIFICATION VIA HUMAN JOINTS USING CONVOLUTIONAL NEURAL NETWORK" (2023). Master's Projects. 1208.