DefinePK

Abstract

Background: Face detection in low-light, low-resolution images remains challenging due to poor contrast, noise, and limited detail. This study proposes a hybrid model using CLAHE with a deep CNN, optimised for robust face detection under such adverse conditions.
Methods: This research proposes a hybrid model integrating CLAHE-based preprocessing with a modified deep CNN for face detection. CLAHE enhances contrast in dark scenes while controlling noise. The Viola-Jones cascade generates candidate face regions, refined by a custom CNN with optimised kernels, batch normalisation, and Spatial Pyramid Pooling for scale invariance. Non-maximum suppression (IoU > 0.5) removes duplicates. The model is trained on WIDER FACE, Dark Face, and additional low-light images, and evaluated on standard benchmarks that focus on low-light and low-resolution accuracy.
Results: The hybrid model achieved a 94.5% face detection rate on extremely low-light, low-resolution images, outperforming YOLOv3, MTCNN, and RetinaFace. On the Dark Face dataset, it showed a higher True Positive Rate and a lower False Positive Rate. The model runs at 12 FPS on CPU, twice as fast as MTCNN, while maintaining superior accuracy. The confusion matrix (Figure 3) shows 94.5% True Positives and 3% False Positives.
Conclusion: CLAHE-enhanced images, combined with a robust modified CNN, enable accurate 94 to 95% face detection in low-light, low-resolution conditions. The hybrid model is well-suited for real-world applications, such as night-time surveillance and mobile devices.