针对当前去雾算法效率不高、细节恢复较差等问题,提出一种基于改进多尺度AOD-Net(All in one dehazing network)的去雾算法。通过增加注意力机制、调整网络结构和改变损失函数这三方面的改进,增强网络的特征提取和恢复能力。模型的第一层增加SPA(Spatial Pyramid Attention,空间金字塔注意力)机制,使网络在特征提取过程中避免冗余信息。将网络改成拉普拉斯金字塔型结构,使模型能够提取不同尺度的特征,保留特征图的高频信息。使用MS-SSIM(Multi Scale Structural Similarity, 多尺度结构相似性)+L1损失函数替换原有的损失函数,提高模型保留结构的能力。实验结果表明,本方法去雾效果更好,细节更丰富。主观上,去雾图像效果优于原网络。客观上,与原网络相比PSNR提升了2.55dB,SSIM值提升了0.04,IE熵值增加了0.18,这证明了算法的出色去雾效果和稳定性。

To address the current issues of inefficient dehazing algorithms and poor detail recovery, we propose an improved multi-scale AOD-Net (All in one dehazing network) algorithm. This algorithm enhances the network"s feature extraction and recovery capabilities through three key improvements: adding an attention mechanism, adjusting the network structure, and modifying the loss function. Specifically, the first layer of the model incorporates the SPA (Spatial Pyramid Attention) mechanism, which enables the network to avoid redundant information during feature extraction. Furthermore, the network structure is modified into a Laplacian pyramid structure, allowing the model to extract features at different scales and preserve high-frequency information in the feature maps. Additionally, the original loss function is replaced with the MS-SSIM (Multi-Scale Structural Similarity) + L1 loss function, thereby enhancing the model"s ability to retain structural information. Experimental results demonstrate that this method achieves better dehazing effects and richer details. Subjectively, the dehazed images exhibit superior quality compared to those produced by the original network. Objectively, compared to the original network, there is a 2.55 dB improvement in PSNR, a 0.04 increase in SSIM value, and a 0.18 increase in IE entropy value, which proves the algorithm"s excellent dehazing effect and stability.

TP391

国家重点研发计划(2019YFB1704502)；国家自然科学(61472173)；江西省研究生创新专项资金项目(yc2023-s995,YJSCX202312)