摘要
针对雾(霾)会显著降低基于图像制导武器的可见光侦测设备成像质量,从而干扰对目标精确识别的问题,提出一种基于条件生成对抗网络的单幅图像去雾算法。在生成器下采样中使用软池化运算,以提高细粒度特征的提取能力;加入全局平均池化层,旨在消除图像边缘的震荡效应,提高去雾图像清晰度;简化判别器结构,优化损失函数权重值确定方法,提升网络模型训练效率。实验结果表明:去雾后的图像清晰锐利,色彩自然,在结构相似性、峰值信噪比和图像信息熵等客观定量指标上优于经典去雾算法,对去雾后图像进行目标检测的平均精度均值提升了4.13%。
Fog(haze) can significantly reduce the imaging quality of visible light detection equipment based on image-guided weapons,thus interfering with the accurate recognition of targets.To solve this problem,a single image defogging algorithm based on conditional generation countermeasure network is proposed.Soft pooling operation is used in the sampling of the generator to improve the extraction ability of fine-grained features.The global average pooling layer is added to eliminate the oscillation effect of image edges and improve the definition of defogged images.The structure of the discriminator is simplified,and the method for determining the weight value of the loss function is optimized to improve the training efficiency of the network model.The experimental results show that the defogged image is clear and sharp with natural color,and it is superior to the classical defogging algorithm in objective quantitative indicators such as structure similarity,peak signal to noise ratio and image information entropy.The average accuracy of target detection in defogged image is improved by 4.13%.
作者
钱坤
李晨瑄
陈美杉
冯宗亮
Qian Kun;Li Chenxuan;Chen Meishan;Feng Zongliang(School of Coastal Defense,Naval Aviation University,Yantai 264000,China;No.32127 Unit of PLA,Dalian 116100,China)
出处
《兵工自动化》
2023年第2期16-23,共8页
Ordnance Industry Automation
基金
装备预研领域基金(6140247030216JB14004)。
关键词
图像去雾
条件生成对抗网络
软池化
损失函数
目标识别
image defogging
conditional generative countermeasure network
soft pooling
loss function
target recognition
作者简介
钱坤(1986-),男,吉林人,博士,从事计算机视觉、图像处理、模式识别研究。E-mail:qk18900992305@163.com。
