Authors :

Asri Abd Rahman and Zaman Ngadi

Address :

Faculty of Computer Science & Information Technology, University of Malaya (UM) Kuala Lumpur, Malaysia.

Department of Computer Science, Faculty of Computing, Universiti Teknologi Malaysia (UTM) Johor, Malaysia

Abstract :

Low-light image enhancement is crucial in computer vision tasks where images suffer from poor visibility, low contrast, and significant noise. The integration of biologically inspired optimization with deep learning models offers a powerful approach for restoring visual quality in such scenarios. Existing methods often rely on fixed enhancement parameters or manually tuned deep networks, which may lead to over-enhancement, color distortion, or poor generalization across varying illumination conditions. To address these limitations, this paper proposes a hybrid framework that combines Owl Eye Optimization (OEO) with Deep Residual Learning (DRL). The OEO algorithm adaptively tunes hyperparameters of the residual network, such as filter size and learning rate, based on perceptual quality metrics, ensuring dynamic adaptability to image complexity. The deep residual network then enhances details, suppresses noise, and improves contrast by utilizing skip connections to maintain structural integrity. The proposed method is applied to low-light image enhancement in surveillance systems and smartphone photography. Experimental evaluations on the LOL dataset show that the proposed model achieves a PSNR of 26.43 dB, SSIM of 0.894, NIQE of 2.51, and LOE of 26.9, outperforming state-of-the-art methods such as RUAS, BSR, and LIMET. The model also demonstrates a convergence rate of 30–35 epochs and an average inference time of 0.052 seconds per image, highlighting its efficiency and robustness in real-time applications.

Keywords :

Low-light enhancement, Deep residual learning, Owl Eye Optimization, Hyperparameter tuning, Image quality, Visual restoration .

DOI :