Child detection in videos using age estimation and convolutional neural networks

• Main Author: Ricafort, David John
• Format: text
• Language: English
• Published: Animo Repository 2018
• Subjects: Neural networks (Computer science); Machine learning
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/5581
• Institution: De La Salle University

Research on facial image age estimation has become popular in recent years due to its possible real-world applications, with different methods being explored for age estimation. Majority of these research used datasets that were taken from controlled environments where the images were high quality, had good lighting conditions and had no occlusions. However, there were also some works that used images from uncontrolled environments and tried to address the issues of datasets that come from an uncontrolled or wild environment. Age estimation can be used in distinguishing between children and adults not just in images, but also in videos. It is important that the age estimation algorithm handle occlusions, poor lighting conditions and low quality inputs because these will be the conditions in real-world scenarios. Recent researches in computer vision show that Convolutional Neural Network (CNN) can handle these conditions, as well as outperform traditional machine learning techniques in image recognition tasks. There are also researches that show how CNN outperform other approaches in face recognition tasks, making CNN an ideal approach for this research. This research detects a child from a video captured from the wild using age estimation and CNN. The best performing CNN model produced in this research achieved an accuracy of 71.13%. It is a 7-layer network that classified inputs into three classes: early childhood, late childhood, and teens/adults.