Improving Personal Protective Equipment Detection in Fabrication Yard Using Image Augmentation with YOLOv7

Accurate object detection is crucial for industrial safety applications such as monitoring the safety of workers in an industrial location. This study examines the impact of applying data augmentation techniques in tandem with the YOLOv7 object detection model to identify three desired classes which are hard hats, safety shoes, and regular shoes. The utilized dataset is 2300 for training images, 400 for validation images, and 600 for test images, covering various scenarios from the industrial environment and general environment. Five experiments were conducted using different combinations of data augmentation methods which are: padding, flipping, resizing, and cropping. The experiment that applied padding, resizing, and flipping to the original 640×640 images achieved the highest mean Average Precision (mAP_0.5) of 0.8597. Another experiment, which combined multiple augmentations across different resolutions, including padding, flipping, resizing, and cropping, produced the second-highest mAP_0.5 of 0.8568. The baseline experiment without augmentation achieved a mAP_0.5 of 0.6412, while resizing images to 320×320 without other augmentations resulted in a slight decrease to 0.6296. An experiment combining resizing, cropping, and flipping on 320×320 images yielded a mAP_0.5 of 0.8323. Based on the results achieved, the augmentation techniques that were applied significantly enhanced the results achieved compared to the non-augmented baseline runs. The findings highlight the effectiveness of data augmentation in improving the generalization capabilities of object detection models by combining diverse augmentation methods like padding, resizing, and flipping proved to be the most effective combination thus suggesting a synergistic effect for the detection of hard hat, safety shoes and shoes.