A Hybrid Approach for Heavily Occluded Face Detection Using Histogram of Oriented Gradients and Deep Learning Models

Abstract

Face detection is a critical component in modern security, surveillance, and human-computer interaction systems, with widespread applications in smartphones, biometric access control, and public monitoring. However, detecting faces with high levels of occlusion, such as those covered by masks, veils, or scarves, remains a significant challenge, as traditional models often fail to generalize under such conditions. This paper presents a hybrid approach that combines traditional handcrafted feature extraction technique called Histogram of Oriented Gradients (HOG) and Canny edge detection with modern deep learning models. The goal is to improve face detection accuracy under occlusions. The proposed method leverages the structural strengths of HOG and edge-based object proposals while exploiting the feature extraction capabilities of Convolutional Neural Networks (CNNs). The effectiveness of the proposed model is assessed using a custom dataset containing 10,000 heavily occluded face images and a subset of the Common Objects in Context (COCO) dataset for non-face samples. The COCO dataset was selected for its variety and realism in background contexts. Experimental evaluations demonstrate significant performance improvements compared to baseline CNN models. Results indicate that DenseNet121 combined with HOG outperforms other counterparts in classification metrics with an F1-score of 87.96% and precision of 88.02%. Enhanced performance is achieved through reduced false positives and improved localization accuracy with the integration of object proposals based on Canny and contour detection. While the proposed method increases inference time from 33.52 to 97.80 ms, it achieves a notable improvement in precision from 80.85% to 88.02% when comparing the baseline DenseNet121 model to its hybrid counterpart. Limitations of the method include higher computational cost and the need for careful tuning of parameters across the edge detection, handcrafted features, and CNN components. These findings highlight the potential of combining handcrafted and learned features for occluded face detection tasks.