ComplexDehazeNet: Complex Multi-Scale Pathway Learning-Enhanced Execution Dehazing Network

In recent years, the application of vision-based systems within the manufacturing industry has increased significantly. However, the presence of hazy images caused by environmental factors, lens contamination, or light scattering often leads to operational inefficiencies and degraded performance in automated systems. To address this issue, in this paper the ComplexDehazeNet is proposed, which is a novel multi-scale and multi-path learning-enhanced dehazing network designed to efficiently restore hazy images. ComplexDehazeNet leverages a hierarchical feature extraction strategy with multi-path learning, enabling the network to capture fine details and global structures simultaneously. By integrating adaptive feature refinement mechanisms and context-aware information processing, the proposed method enhances both local contrast and overall image clarity. Additionally, the proposed model incorporates an advanced attention module that dynamically adjusts feature importance, ensuring effective dehazing across regions with uneven haze distributions. The use of residual learning and adaptive skip connections further improves the training stability and visual reconstruction quality. The proposed method achieves a PSNR of 44.46 dB and an SSIM of 0.9974 on the RESIDE-Indoor dataset, outperforming existing state-of-the-art dehazing models. The code is available at https://github.com/DavidcOfficial/ComplexDehazeNet/.