Publication: Lightweight face detection model
Loading...
Date
2024-07
Authors
Lam, Zi Yao
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Nowadays, object detection technologies have become popular among various industries fields due to its high efficiencies and durability in performing industrial
operations. This technology can be utilized by a wide range of individuals such as performing face recognition using electronic devices on people. Face recognition is a biometric technology that match human face from image or video frame to a database of individual with certified identity. Currently there are an increase in the demand of lightweight face detection model to be implemented on smaller device. However, most of the CNN based models that exist within the present time are computationally heavy and required many resources with time to train. This project aims to modify the current model of Ultralytics Yolov8 architecture to improve its performance and reduce its computational cost such that the model is more compatible with lightweight devices. The modification can be done by changing the convolution operation in the architecture blocks of Yolov8. The modified model will then be trained and evaluated based on their performance parameters such as FLOPs, F1 score and mAP. The results shows that the n model utilizing DWConv has the least parameters, while having a
precision of 84.61% and accuracy of 74.13%. The model which has the best effectiveness is m model utilizing GhostConv as it has mAP50 of 80.97% and mAP50 95 of 58.65%. Through the results, model utilizing DWConv has the least FLOPs model utilizing GhostConv tends to have better mAP with increasing parameters.