MS in Computer Science (M.S.C.S.)
Degree Granting Department
Engineering Computer Science
Dmitry Goldgof, Ph.D.
Yu Sun, Ph.D.
Shaun Canavan, Ph.D.
bias, dropbox, mish, mosaic, YOLOv3
There are many face detection classification models available for download and use in the modern technological world. Based in the field of deep neural networks, these off-the-shelf solutions are generally inadequate to solve real world challenges. This work presents how current approaches biased towards detecting adult human faces must be modified in order to better accommodate face detection of the neonate in a NICU setting.
YOLO is a powerful object detection algorithm. Due to optimizations such as Cross mini-batch Normalization, Modified Spatial Attention Modules, Modified Path Aggregation Networks, Self-Adversarial Training, Mosaic Data Augmentation, DropBox Regularization, Multi-Input Weighted Residual Connections and Mish-activation functions, in its most recent form, YOLOv5 will be lightweight enough to be integrated into the clinical setting. As YOLO has always been a front runner in terms of computation speed, with the optimizations above YOLOv5 is able to make real-time detection. With appropriate training, YOLOv5 will be able to accurately detect the face of the neonate and overcome bias, occlusion and extremely difficult background context that normally cause inaccurate results with other object detection methods.
Scholar Commons Citation
Hausmann, Jacqueline, "The Efficiency and Accuracy of YOLO for Neonate Face Detection in the Clinical Setting" (2020). USF Tampa Graduate Theses and Dissertations.