To address issues such as incomplete dehazing and distortion in image dehazing, a multi-depth adaptive fusion dehazing generation network is proposed using generative adversarial training. Firitly, the network utilizes U-Net++ architecture and haze perception units to learn the haze features. Secondly, a vector mixed attention module is proposed to expand the bottom-layer information and supplement details in the dehazed image. Thirdly, adaptive weights are constructed from partial and global dimensions to select features at different depths and improve the utilization of effective information in the network. Finally, a mixed loss is employed to ensure the quality of the generated image, and Wasserstein distance is introduced into the adversarial loss. To validate the effectiveness of the proposed algorithm, objective quantitative comparisons are conducted with 10 popular dehazing algorithms on the RESIDE and Haze4k datasets, followed by subjective evaluations on real images. The experimental results demonstrate that the proposed algorithm achieves a PSNR of 36.20 dB and SSIM of 0.988 4 on the SOTS Outdoor validation set, showing superior dehazing performance.
首先,采用深度可分离卷积和不同卷积混合的方式,减少网络参数量,扩展网络的深度。其次,为了更加聚焦于有雾区域,提出全局感知模块,在水平和竖直方向上进行全局池化,通过拆分学习再合并的方式获取全局感知特征图。然后,用全连接的方式代替卷积,实现通道数的改变,相较于卷积只关注局部信息,全连接的方式能获取图像的整体信息,并采用反瓶颈结构,在中间层4倍扩张向量,引入一个比输入和输出更高的维度,从而增加网络的深度和复杂度,增强模型的表达能力。考虑到网络在训练的过程中,易发生过拟合和网络退化的问题,提出层缩放和随机子路径删除的策略来解决。最后,采用跳跃连接的方式,增加输入特征和输出特征的交互,从而使网络更容易学习到雾霾的特征,提高图像的保真度。为了保证梯度传输过程的稳定性,在非线性激活函数的选取上,采用GELU(Gaussian error linear unit)激活函数代替常用的Relu(Recitified linear unit)激活函数。相较于Relu激活函数,GELU激活函数在零处的梯度更加平滑。在归一化函数的选取上,采用LN(Layer normalitzation)层代替BN(Batch normalitzation)层,由于BN层是从Batch维度进行标准化,所以当Batch很小时,效果并理想,而LN层是从通道维度进行标准化,并不受Batch size的影响。
为了验证本文算法的有效性,在合成图片和真实图片上进行实验,采用RESIDE[17]和Haze4k[18]两个公共数据集。在RESIDE数据集上进行实验时,分室内和室外两组进行。室内训练采用室内合成数据集(Indoor training set,ITS)中的13 990幅图片,室外训练采用室外合成数据集(Outdoor training set,OTS)中不同场景的8 970幅图像,并在合成客观数据集(Synthetic objective testing set,SOTS)上进行测试。在Haze4k数据集上进行实验时,采用训练集和测试集中的全部图片。
在图像质量评价中,分为合成数据和真实数据集两个部分。在合成数据中,将本文的算法和其他10种主流算法(DCP[2]、CAP[3]、AOD-Net[7]、Dehazed[6]、FD-GAN[10]、HIDE-GAN[11]、GridDehazed[8]、FFA-Net[12]、SADNet[19]、GANID[20])从客观层面进行对比,并挑选其中6种算法进行可视化的主观分析。在真实数据集中由于缺少无雾的对比图,同样采用主观分析的方式。在客观对比时,采用峰值信噪比(Peak signal to noise ratio,PSNR)、结构相似性(Structural similarity,SSIM)。峰值信噪比PSNR通过计算两幅图像各个像素的差异来实现对图像的评价,其值越大代表图像质量越高。结构相似性SSIM通过分析图片亮度、对比度和结构3个因素来判断图片质量的好坏,其设计更符合人眼视觉感知系统的特点,取值范围为0~1,相似度越大,图片质量越高。
YaoD N L, BadeA, ZolkiflyI A, et al. A naive but effective post-processing approach for dark channel prior (DCP)[C]∥Data Science and Emerging Technologies: Proceedings of DaSET 2022, Petaling Jaya, Malaysia, 2022: 67-76.
[2]
HeK M, SunJ, TangX O. Single image haze removal using dark channel prior[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33(12): 2341-2353.
[3]
RaananF. Dehazing using color-lines[J]. ACM Transactions on Graphics, 2014, 34(1): 1-14.
[4]
TanR T. Visibility in bad weather from a single image[C]∥Proceedings of the 2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2008), Anchorage, USA, 2008: 1-8.
[5]
ZhuQ S, MaiJ M, ShaoL. A fast single image haze removal algorithm using color attenuation prior[J]. 2015, 24(11): 3522-3533.
[6]
CaiB, XuX, JiaK, et al. Dehazenet: an end-to-end system for single image haze removal[J]. IEEE Transactions on Image Processing, 2016, 25(11): 5187-5198.
[7]
LiB, PengX, WangZ, et al. Aod-net: all-in-One dehazing network[C]∥proceedings of the 2017 IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 2017: 4780-4788.
[8]
LiuX, MaY, ShiZ, et al. Grid dehaze net: attention-based multi-scale network for image dehazing[C]∥Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea, 2019: 7313-7322.
[9]
QuY, ChenY, HuangJ, et al. Enhanced pix2pix dehazing network[C]∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, 2019: 8152-8160.
[10]
DongY, LiuY, ZhangH, et al. Fd-gan: generative adversarial networks with fusion-discriminator for single image dehazing[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(7): 10729-10736.
[11]
MehtaA, SinhaH, NarangP, et al. HIDeGan: a hyperspectral-guided image dehazing GAN[J]. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Seattle, USA, 2020: 846-856.
[12]
QinX, WangZ L, BaiY C, et al. FFA-Net: feature fusion attention network for single image dehazing[J/OL]. [2023-07-22].
[13]
YangX, YanJ C, MingQ, et al. Rethinking rotated object detection with gaussian wasserstein distance loss[J/OL]. [2023-07-23].
[14]
IsolaP, ZhuJ Y, ZhouT, et al. Image-to-image translation with conditional adversarial networks[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 5967-5976.
[15]
ZhaoC, ShuaiR, MaL, et al. Segmentation of skin lesions image based on U-Net++[J]. Multimedia Tools and Applications, 2022, 81(6): 8691-8717.
[16]
FuH X, SongG Q, WangY C. Improved YOLOv4 marine target detection combined with CBAM[J]. Symmetry, 2021, 13(4): No.623.
[17]
LiB, RenW, FuD, et al. Benchmarking single-image dehazing and beyond[J]. IEEE Transactions on Image Processing, 2018, 28(1): 492-505.
[18]
LiuY, ZhuL, PeiS D, et al. From synthetic to real: image dehazing collaborating with unlabeled real data[J/OL]. [2023-07-23].
[19]
SunZ, ZhangY, BaoF, et al. SADnet: semi-supervised single image dehazing method based on an attention mechanism[J]. ACM Transactions on Multimedia Computing, Communications, and Applications, 2022, 18(2): 1-23.
[20]
ManuC M, SreeniK G. GANID: a novel generative adversarial network for image dehazing[J]. The Visual Computer, 2022, 39: 2923-3936.