SINGLE IMAGE DEHAZING BASED ON IMPROVED AOD-NET

Hazy images often suffer from low quality when processed by traditional defogging algorithms, which fail to effectively remove haze. To address this issue, this article propose a novel single-image dehazing model based on the AOD-Net architecture. The model leverages depthwise separable convolutions to construct a lightweight deep neural network. Additionally, this article introduces a new conditional convolution and attention mechanism to enhance feature extraction, thereby improving the network’s ability to capture global information from hazy images. To optimize model performance, this article train the proposed model on the NYU dataset and conduct extensive experiments on the same dataset. The dehazing effectiveness is evaluated using full-reference image quality assessment metrics. Experimental results demonstrate that the improved model achieves higher accuracy in dehazing quality compared to existing methods. Furthermore, the incorporation of the new feature extraction module and attention mechanism significantly enhances performance in haze removal, color restoration, and detail preservation, outperforming the original AOD-Net and other traditional approaches.The application feasibility of this technique is extensive: In the domain of autonomous driving, it can enhance the target detection accuracy of on-board cameras in foggy weather; in remote sensing monitoring, it facilitates satellites and unmanned aerial vehicles to obtain clearer surface information; in the area of security surveillance, it can strengthen the reliability of video analysis in low-visibility circumstances. Additionally, the lightweight design of the model can be adapted to edge computing devices, providing technical support for real-time defogging and possessing significant engineering application value and commercial potential.