HybridEnhancementNet: A Novel Deep Learning Architecture for Low-Light Image Enhancement through Model Fusion

Abstract: Deep learning models replicate key aspects of human cognition, enabling powerful capabilities in classification, recognition, and generation. In areas like vehicle autonomy, computer vision and artificial intelligence (AI) play a critical role—especially in low-light conditions where enhanced visual input can improve safety and performance.

A promising strategy in model development is combining complementary architectures to leverage their individual strengths. This work introduces HybridEnhancementNet (HybridNet), a novel low-light image enhancement framework that fuses the convolutional enhancement and ReLU activation function of the Zero-Reference Deep Curve Estimation (Zero-DCE) model with the feature concatenation and skip-connections of the Color and Intensity Decoupling Network (CIDNet) model. The model is trained on the Low-Light (LOL) dataset, consisting of 500 image pairs captured in low-light and well-lit conditions. Experimental results demonstrate that HybridEnhancementNet produces enhanced images with improved visibility while preserving detail and smooth contrast transitions. The model showed robust generalization, delivering strong performance on the 6,000-image DarkFace low-light face-detection dataset, despite having no prior exposure to it. These findings highlight the potential of hybrid deep learning architectures for real-world computer vision applications, including autonomous navigation under challenging lighting conditions. Future work should further explore model fusion strategies to optimize performance and robustness across diverse visual environments.