YOLO Method Analysis and Comparison for Real-Time Human Face Detection

Face detection plays a huge role in the fields of computer vision and pattern recognition. YOLOv3 is a fast single-stage detector method. However, if consider real applications like mobile devices which have limited memory and computation. YOLOv3 with darknet-53 still has a complex architecture, that can have an impact on computing costs. This study specifically focuses on the comparative analysis of the YOLOv3 method on the problem of human face detection by proposing YOLOv3-Tiny that adopts YOLOv3. However, with a reduced architecture to increase detection speed and make the method lightweight, so as compatible with the needs of low-power devices with limited memory and computing. Both methods were trained with human facial data. The results of the training with 50 epochs YOLOv3 outperformed YOLOv3-Tiny by a difference of 7% precision, 3% recall, and 10% mAP. However, in terms of time consumption during the training process and the size of the resulting weight. YOLOv3-Tiny shows much better results, which only takes 39 minutes, and weighs size only 16.6 MB, while YOLOv3 takes 5 hours 9 minutes with weighs 117 MB. Both methods were then tested on RGB and Grayscale images. Tests are carried out based on the distance of the face and the number of faces. The test results of the YOLOv3 are very good at predicting close faces with the highest value of 95% and getting the same results on many faces with an average value of 95% accuracy. YOLOv3-Tiny achieves lower accuracy than YOLOv3 on both image types. The highest result of YOLOv3-Tiny is close face with 87% accuracy. However, YOLOv3-Tiny excels in detection speed. The final test is based on the speed of CCTV video inference. YOLOv3-Tiny significantly outperformed YOLOv3 with detection speeds of 104.1 FPS on the GPU and 8.7 FPS on the CPU. Meanwhile, YOLOv3 is only able to detect faces at a speed of 30.2 FPS on the GPU and 1.6 FPS on the CPU.