A Composite Noise Removal Network Based on Multi-domain Adaptation

the limitation of conventional singlescene image denoising algorithms in filtering mixed environmental disturbances, and recognizing the drawbacks of cascaded image enhancement algorithms, which have poor realtime performance and high computational demands, The composite weather adaptive denoising network (CWADN) is proposed. A Cascade Hourglass Feature Extraction Network is constructed with a visual attention mechanism to extract characteristics of rain, fog, and low-light noise from authentic natural images. These features are then transferred from their original real distribution domain to a synthetic distribution domain using a deep residual convolutional neural network. The generator and style encoder of the adversarial network work together to adaptively remove the transferred noise through a combination of supervised and unsupervised training, this approach achieves adaptive denoising capabilities tailored to complex natural environmental noise. Experimental results demonstrate that the proposed denoising network yields a high signal-to-noise ratio while maintaining excellent image fidelity. It effectively prevents image distortion, particularly in critical target areas. Additionally, it adapts to various types of mixed noise, making it a valuable tool for preprocessing images in advanced machine vision algorithms such as target recognition and tracking.