Evaluating Low-Light Image Enhancement for Object Detection

Background: Low-light image enhancement is often applied before object detection on the assumption that improved visibility also improves machine perception. This study examines that assumption while separating the immediate effect of enhancement from detector adaptation. Methods: Original ExDark images were compared with Gamma correction, CLAHE, Zero-DCE, and Retinexformer using YOLO11n. Two protocols were evaluated: inference-only enhancement, in which a detector trained on original images was tested across all image domains, and matched training, in which separate detectors were trained and tested within each domain. Matched training was repeated using three random seeds. Performance was assessed with [email protected], [email protected]:0.95, precision, recall, F1-score, paired bootstrap confidence intervals, NIQE, and BRISQUE. Results: Inference-only enhancement reduced detection performance for all four enhancement methods. Matched training recovered most of this loss. Across three seeds, the original images achieved the highest mean [email protected] (0.5563 ± 0.0056), followed closely by CLAHE, Gamma, and Zero-DCE. Retinexformer achieved the best NIQE and BRISQUE scores but the lowest matched detection performance (0.5194 ± 0.0054). Conclusion: Train-test distribution mismatch explains much of the apparent degradation caused by low-light enhancement. None of the tested enhancement methods consistently outperformed the unenhanced baseline, and improved perceptual quality did not reliably predict better object detection.