A Systematic Review on A Knowledge-Distilled Multi-Task Learning Framework for Real Time Low-Light Face Detection

Abstract

Real-time face detection in low-light conditions is critical for modern surveillance but faces significant challenges due to poor visibility and high noise. Traditional sequential pipelines often suffer from task discrepancy, where image enhancement optimized for visual metrics fails to improve downstream detection accuracy. This systematic review evaluates the integration of Knowledge Distillation (KD) and Multi-Task Learning (MTL) to address these limitations. The primary purpose is to identify an efficient, end-to-end solution that can run on low computational power. Following a Systematic Literature Review (SLR) methodology, this study analyzes peer-reviewed research from 2020 onwards, focusing on Teacher-Student architectures and joint optimization strategies. The review identifies a critical gap in existing literature: the lack of unified frameworks that balance visual restoration with functional detection performance. Key findings highlight how KD transfers robust feature extraction from complex models to lightweight students, while MTL utilizes hybrid loss functions to ensure enhancement is directly guided by detection requirements. Synthesis of the literature demonstrates that integrated KD-MTL frameworks significantly outperform decoupled methods in both inference speed (FPS) and mean Average Precision (mAP) on datasets like DARKFACE. The paper concludes that the integration of Knowledge Distillation and Multi-Task Learning successfully addresses the disparity between image quality restoration and face detection while maintaining low computational requirements, providing a practical solution for applications in real-world surveillance and night photography.