基於色調映射與模型可解釋性技術的人臉偵測優化

Abstract

邊緣攝影機在極端背光與低光環境下，因對比失衡與雜訊升高，常導致人臉偵測表現顯著退化。本研究以全域與區域色調映射為核心，結合輕量化偵測器進行系統性評估與消融，聚焦於「前端影像增益」與「小樣本重新訓練」的相對效益與互補效應。結果顯示，在背光與低光影像集中，最佳組合可將檢測精度由 11.6% 提升至 50.7 % ，並明顯改善困難區域的人臉可見度與穩定性。基於此結論，我們提出適用於資源受限情境的實作指引，說明前端增益與輕量偵測的搭配原則與取捨，提供可部署方案，並為後續自適應色調映射與輕量偵測器的協同設計奠定基礎。

Edge cameras operating under extreme backlit and low-light conditions often suffer substantial degradation in face detection due to contrast imbalance and elevated noise. Centered on global and local tone-mapping strategies, this study conducts a systematic evaluation and ablation in combination with lightweight detectors, focusing on the relative benefits and complementary effects of front-end image enhancement and small-sample retraining. Results show that, on backlit and low-light image sets, the best configuration increases detection accuracy from 11.6% to 50.7% and markedly improves face visibility and stability in challenging regions. Building on these findings, we provide implementation guidelines for resource-constrained settings, outlining the pairing principles and trade-offs between front-end enhancement and lightweight detection, offering deployable solutions and laying the groundwork for subsequent co-design of adaptive tone mapping and lightweight detectors.