Robust low-light object detection and classification via hybrid feature fusion and synthetic data augmentation
Abstract Robust visual navigation in low-light environments remains a fundamental challenge for autonomous systems due to severe illumination degradation, low signal-to-noise ratio, reduced texture information, and loss of discriminative visual cues. These primary conditions significantly impact the reliability of conventional handcrafted descriptors to detect objects on training on well-illuminated datasets through deep learning models. To improve low-light image object detection and classification by overcoming above limitation, we proposed a low-light visual recognition framework that combines image enhancement, complementary handcrafted and pretrained deep feature representations, dimensionality reduction, and synthetic low-light data augmentation. In the pre-processing stage, we had applied Contrast Limited Adaptive Histogram Equalization (CLAHE), followed by gamma correction, to enhance local contrast while preserving structural information. To improve generalization across varying illumination levels, a pseudo low-light image synthesis pipeline was introduced, combining gamma-based exposure reduction, Poisson photon noise modeling, and intensity bias simulation. The resulting pseudo low-light images provide realistic illumination degradations for data augmentation. Feature extraction was done using Histogram of Oriented Gradients (HOG), which captures local edge and gradient structures that are comparatively robust to intensity variations, followed by deep convolutional embeddings that provide high-level semantic representations. For image classification, HOG descriptors and pre-trained features of ResNet-18 model was evaluated both individually and as a fused representation with PCA-based dimensionality reduction, using Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP) classifiers. Separately, for object detection, synthetic low-light images generated from COCO images were used to augment ExDark training images, and their effect is evaluated with YOLOv8 and YOLOv11 object detection models. Extensive experiments were performed on the ExDark benchmark dataset to evaluate the proposed hybrid feature fusion strategy in terms of feature discriminability, detection accuracy, recall, and robustness across varying illumination conditions, in comparison with conventional handcrafted and deep feature baselines. The results show that complementary feature representations and synthetic low-light augmentation improve classification and detection performance under challenging low-light conditions, supporting improved robustness to illumination variation of low-light images.
Authors
- Shawn Sethi (ORCID: https://orcid.org/0000-0002-8734-103X)
- Rayappa David Amar Raj (ORCID: https://orcid.org/0000-0002-5888-5513)
- Akshaya Vasudevan (ORCID: https://orcid.org/0000-0002-2733-2115)
- Srija Manimaran (ORCID: https://orcid.org/0000-0001-6918-4678)
- Karishini S
- Gowri J S
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1038/s41598-026-78730-9
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00