基于改进AOD-Net网络模型的车载图像去雾方法  

Improved AOD net network model for vehicle image defogging

作　　者：景嘉宝 王正家[1,2] 何涛 翟海祥[1,2] JING Jiabao;WANG Zhengjia;HE Tao;ZHAI Haixiang(School of mechanical engineering,Hubei University of Technology,Wuhan 430068,China;Key Laboratory of·modern manufacturing quality engineering,Wuhan 430068,China)

机构地区：[1]湖北工业大学机械工程学院,武汉430068 [2]现代制造质量工程湖北省重点实验室,武汉430068

出　　处：《激光杂志》2023年第2期83-90,共8页Laser Journal

基　　金：国家自然科学基金项目(No.51275158)。

摘　　要：基于现有去雾算法存在去雾效果不佳和去雾效率低等问题,提出了一种改进AOD-Net网络模型。首先对输入图像进行随机噪声添加,提高图像模型去雾鲁棒性。接着对不同尺度的卷积核进行多线程处理,同时将图像中的特征信息提取,然后利用注意力机制进行权重分配,采集图像中的纹理信息和细腻化特征信息,提升图像的质量。最后对提取的特征信息利用AOD-Net模型的前两层卷积进行二次特征提取,估计出联合参数,输出去雾后的图像。实验结果表明,采用本算法得到的第一组和第二组图像峰值信噪比分别为20.05和16.92,结构相似性分别为0.85和0.83,IE熵值分别为7.48和7.75,各项数值均有提升,图像去雾效果更好,证明了本算法的有效性。Based on the problems of poor dehazing effect and low dehazing efficiency in existing dehazing algorithms, this paper proposes an improved AOD-Net network model. Firstly, random noise is added to the input image to improve the dehazing robustness of the image model. Then, multi-thread processing is performed on the convolution kernels of different scales, and the feature information in the image is extracted at the same time, and then the attention mechanism is used for weight distribution, and the texture information and refined feature information in the image are collected to improve the quality of the image. Finally, the extracted feature information is extracted by the first two layers of convolution of the AOD-Net model, and the joint parameters are estimated, and the dehazed image is output.The experimental results show that the peak signal-to-noise ratios of the first and second groups of obtained by this algorithm are 20. 05 and 16. 92, respectively, the structural similarity is 0. 85 and 0. 83, and the IE entropy values are 7. 48 and 7. 75, respectively. The dehazing effect is better, which proves the effectiveness of the algorithm in this paper.

关 键 词：图像去雾 暗通道先验 卷积神经网络 AOD-Net模型 端到端 

分 类 号：TN249[电子电信—物理电子学]