基于改进CycleGAN模型的图像去雾方法  

Image Defogging Method Based on Improved CycleGAN Model

作　　者：王旭光 张崇 田珊珊 白康 WANG Xu-guang;ZHANG Chong;TIAN Shan-shan;BAI Kang(Hebei Technology Innovation Center of Simulation&Optimized Control for Power Generation,North China Electric Power University,Baoding Hebei 071003,China)

机构地区：[1]华北电力大学河北省发电过程仿真与优化控制工程技术研究中心,河北保定071003

出　　处：《计算机仿真》2024年第7期250-257,410,共9页Computer Simulation

基　　金：国家科学自然基金项目(62076093);河北省省级科技计划资助(22567643H)。

摘　　要：雾天条件下拍摄的图像不够清晰,严重影响到后续计算机视觉任务的完成质量,因此,对雾天拍摄图像进行清晰化处理具有实际意义。现有CycleGAN去雾模型未能充分利用在编码和解码过程中的特征信息,得到的去雾图像质量不够高。生成器通过设计的特征融合模块引入多尺度信息,鉴别器采用网络共享的双鉴别器策略,损失函数使用两种对抗损失并结合改进的循环一致性损失,增强模型去雾性能和图像质量。对比实验和消融实验验证了改进模型的性能。The image capture under foggy condi tions is not clear enough,which seriously affects the quality of subsequent computer vision tasks.Therefore,it is of practical significance to clear the image taken on hazy days.The existing CycleGAN dehazing model fails to make full use of the feature information in the pr ocess of encoding and decoding,and the quality of the dehazing image obtained is not high enough.In the method,the generator introduces multi-scale information through the designed feat ure fusion module,the discriminator adopts the network shared double discriminator strategy,and the loss function uses two kinds of adversarial loss combined with the improved cycle consistency loss to enhance the model's dehazing performance and image quality.Comparisonex periment and ablation experiment verify the performance of the improved model.

关 键 词：多尺度融合 双鉴别器 图像去雾 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]