基于生成对抗网络的图像去雾算法研究

Image Dehazing Algorithm Based on Generative Adversarial Networks

摘要: 为解决目前基于先验知识与深度学习的图像去雾算法底层网络感受野较小，不能有效捕捉远距离的依赖信息、对图像全局特征利用不充分导致复原的图像整体恢复效果不佳、细节恢复不完整、图像颜色失真等问题，本文提出了基于生成对抗网络改进的图像去雾算法。生成器模块首先通过DenseNet网络进行初步特征提取，其次设计ASPP-Transformer模块，通过ASPP模块特征进行多尺度特征提取，再通过Transformer模块捕捉图像的全局特征依赖关系，实现对提取的多尺度局部特征信息与全局信息特征信息的有效融合，提升了生成器在复杂雾气场景中的全局一致性表现，最后通过解码器对融合后的特征进行重构，得到无雾图像。实验表明，通过本文算法恢复的无雾图像在主观质量与客观指标上都取得较好的评价结果。

Abstract: To address issues in current image dehazing algorithms based on prior knowledge and deep learning—such as limited receptive fields in underlying networks, insufficient capture of long-range dependencies, inadequate utilization of global features, incomplete restoration of image details, and color distortion—this paper proposes a novel image dehazing algorithm based on Generative Adversarial Networks (GAN). The generator module first performs feature extraction using a DenseNet network, followed by an ASPP-Transformer module that integrates multi-scale feature extraction through the ASPP module and captures global feature dependencies with the Transformer module. This enables effective fusion of multi-scale local and global information, enhancing the generator’s global consistency in complex hazy scenes. Finally, a decoder reconstructs the fused features to produce a dehazed image. Experimental results demonstrate that the proposed algorithm achieves superior subjective quality and objective performance in restoring dehazed images.