基于生成对抗网络的喷雾液滴运动模糊恢复

Spray droplet motion fuzzy recovery based on generating adversarial networks

针对喷雾液滴的显微拍摄快速测量方法中造成的运动模糊问题,提出了基于改进的DeblurGAN网络的去模糊方法,在保证模型精度的情况下精简模型。特征提取模块使用深度可分离卷积取代常规的卷积,且扩大卷积核为5×5;同时引进改进AECA注意力机制让网络更加关注于液滴图片中的重要特征;构建AECA-FPN特征金字塔结构,融合多尺度的特征。对液滴图像做运动模糊仿真,构建数据集,训练网络。实验结果表明,算法在峰值信噪比(PSNR)和结构相似性(SSIM)指标上对比原网络与维纳滤波法都获得了较大的提升。对拍摄的液滴模糊图像进行识别恢复,原模糊图像的拉普拉斯方差平均值从12.51上升到25.97,对显微液滴图像的运动模糊有较好的恢复效果。

Aiming at the motion blur problem caused by the rapid measurement method of droplet microscopy, the deblurGAN adversarial generation network based on improved DeblurGAN is proposed. Streamline models while ensuring model accuracy.and the depth separable convolution is used in the feature extraction module to replace the conventional convolution, and the convolutional kernel is enlarged to 5×5;at the same time, the improved AECA attention mechanism is introduced to make the network pay more attention to the important features in the droplet picture;the AECA-FPN feature pyramid backbone network is constructed, and the multi-scale features are fused to improve the network accuracy;the motion blur simulation of the droplet image is constructed, the data set is constructed, and the network is trained. Simulation experiments show that the proposed algorithm has greatly improved the original network and Wiener filtering method on the PSNR and SSIM indicators. The droplet blurred images were recognized and restored, and the average of Laplace variances of the original blurred images rose from 12.51 to 25.97. The proposed algorithm has a good recovery effect on the motion blur of the microscopic droplet image.