Optimizing facial feature extraction and localization using YOLOv5: An empirical analysis of backbone architectures with data augmentation for precise facial region detection

The task of object detection in computer vision revolves around the identification of objects within images or videos. A specific subtask within object detection is face detection, which focuses on detecting human faces. Within the realm of face detection, an important research area is facial feature detection, which has diverse applications ranging from facial recognition to emotion detection and facial expression analysis. The crucial step in facial feature detection is the identification and localization of key facial features such as the eyes, eyebrows, nose, mouth, and chin, which can also be called facial region detection. Face region detection can be done in two ways: landmark detection and Bounding box- based detection. Bounding boxes offer computational benefits such as increased speed and efficiency. They are preferable when the objective is to accurately detect and locate the presence of an object or face in an image or video frame. Although most of the existing algorithms for facial feature detection based on bounding box predictions typically treat the eyes as a single entity, our approach using YOLOv5 addresses the separation of left and right eye detection. In this research study, we conducted experiments using YOLOv5, which provides bounding box predictions. We used a subset of LFW (Labelled Faces in the Wild) Dataset which we augmented using GFP-GAN, Gaussian Noise, Image Sharpening, and CLAHE. We explored the effectiveness of different backbone architectures when applied to YOLOv5 for the task of facial region detection. We evaluated three popular backbone networks: EfficientNet-b0, GhostNet, and CSP-Darknet53. Our objective was to identify the most suitable backbone architecture that yields accurate detection of facial features, including the left eye, right eye, nose, and lips. Our experiments show that when GhostNet is used as a backbone in the YOLOv5 architecture, it produces superior results for the detection and classification of features as compared to the other backbones. We present a detailed evaluation of our findings, including discussions of the experimental results using different IOU thresholds and backbone combinations. Our proposed methodology and findings make valuable contributions to the field of facial feature extraction and provide meaningful insights into the potential and performance of YOLOv5 for detecting and localizing key facial elements.