Adaptive Low-Light Image Enhancement Selection Using Random Forest: Validation on ExDark Dataset with BRISQUE Quality Assessment

Abstract

Low-light image enhancement remains challenging for resource-constrained edge devices where deep learning models are often too computationally expensive. This research presents an adaptive selection framework that learns to choose between two complementary enhancement strategies—a lightweight logarithmic boost and a sophisticated realistic fusion method—based on image characteristics. A Random Forest classifier trained on 7 PCA-reduced features from 1,385 images achieves 86.1% accuracy for method selection, with 5-fold cross-validation confirming robust performance (84.3% ± 1.2%). This method is tested on ExDark dataset containing 1200 images from 12 classes. Object detection uses blur and image quality metrics (BRISQUE, NIQE, PIQE) along with YOLOv8. The results show clear differences, with a 2.1% improvement in detection rate (p = 0.017), a 10% improvement in BRISQUE quality (p < 0.001), and a 26.7% increase in brightness (p < 0.001). When combined with other methods, it leaves behind Histogram Smoothing (61.5%) and CLAHE (68.0%) and matches Gamma Correction (70.5%), but BRISQUE scores (21.93 and 26.78) make images look more natural. On closer inspection, certain categories such as bicycles show significant increases, but boats, bottles, and dining tables do not receive better detection scores. It shows clear scenarios where this approach works well and where it doesn’t. Additionally, this method takes up almost no space at just 1.86 MB and runs at 18.6 FPS on an average CPU without requiring any special graphics hardware. This makes it handy for edge devices like cameras or phones, especially in low light where sharpening edges and retaining detail is critical. The method is simple, requires little space and offers better results for certain classes without requiring large resources.