Face Mask Wearing Detection Using Yolo-v4 Algorithm

Lim, Jie Qian (2023) Face Mask Wearing Detection Using Yolo-v4 Algorithm. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text -Lim Jie Qian.pdf Restricted to Registered users only Download (7MB)

Abstract

There are many real-time approaches to performing face mask wearing detection. The most famous method for real-time face mask wearing detection will be the CNN approach. CNN approaches having a very big family, which it can separate into single-stage detectors and doublestage detectors. The two-stage object detector such as R-CNN, Fast R-CNN, Faster R-CNN, and Mask R-CNN normally has the highest accuracy in object detection, but the detection will be time-consuming and unsuitable for a real-time face mask wearing detector, where the FPS is lower than the single-stage object detection. It is because it had multiple inference steps per second. The single-stage detectors such as SSD and YOLO only need a single pass through the neural network to perform the object detection, so it is suitable for real-time face mask wearing detection. Although SSD has better accuracy than the YOLO method, it has trouble recognising objects of smaller sizes. For research, the YOLO family, especially YOLO-v4, can produce very high accuracy for object detection at a very fast response time. Therefore, YOLO-v4 will be further studied because of its real-time and stable accuracy performance. YOLO-v4 is a popular object detection algorithm, widely used for its high precision and real-time processing capabilities. However, due to its complex neural network architecture and the substantial computations required for object detection, YOLO-v4 requires powerful GPUs for efficient operation. Google Colab, an online GPU service, can provide access to highperformance GPUs, despite limitations such as time constraints, memory and storage space, and network latency. While YOLOv4 exhibits high precision and real-time processing capabilities, detecting small objects and complex scenes in images or videos remains a challenge. Therefore, this algorithm was trained and evaluated using the Kaggle face mask detection dataset. Some image pre-processing and hyperparameter are fine-tuned to optimize the performance of the algorithm, such as learning rate, input size, saturation, exposure, and some augmentations. The best configuration was selected, and two layers are added to the PAN algorithm to further improve it. After this series of optimizations, the accuracy significantly increased, but the speed decreased due to the added complexity of the layers, more parameters to learn, and more computations to perform. Finally, the improved YOLO-v4 achieved an IoU of 76.32%,