基于cGAN的下采样LG谱图像优化重建

Optimal image reconstruction from down-sampled LG spectrum based on cGAN

对于复杂图像的拉盖尔高斯(Laguerre-Gaussian,LG)谱成像,因为满足奈奎斯特采样率的高阶LG模式系数无法测得,重建图像的失真不可避免,而神经网络算法通过先验学习,可以对失真图像实现较为清晰的复原.提出基于条件生成对抗网络(Conditional Generative Adversarial Nets,cGAN)的图像优化重建方法,在处理下采样的LG谱单像素成像和旋转运动模糊图像中均取得了较好的效果.在1.87%的LG谱采样率下,该方法能将Kaggle数据集人像二值图像的结构相似性(Structural Similarity,SSIM)指数提升至0.8以上,和经典图像去噪算法相比有显著提升.

For Laguerre-Gaussian(LG)spectral imaging under Nyquist sampling rate,the reconstructed images are generally distorted because it is difficult to measure the higher-order LG mode coefficients.The neural network algorithm can be used to restore these images through prior learning.In this paper,we propose an optimal image reconstruction method based on Conditional Generative Adversarial Nets(cGAN),which works well in down-sampled LG spectral single-pixel imaging and rotational motion blur recovery in LG spectral domain.We use the portrait binary images from Kaggle dataset as an example.At a sampling rate of 1.87%,the structural similarity(SSIM)index by using our method reaches 0.8 and above,which is significantly improved comparing with classical image denoising algorithms.