光学
物理
可分离空间
迭代重建
散射
图像处理
图像质量
图像(数学)
光散射
计算机科学
空间频率
逆散射问题
前向散射
点扩散函数
傅里叶变换
反射(计算机编程)
几何光学
相位恢复
衍射
图像复原
全息术
物理光学
折射率
卷积(计算机科学)
切趾
激光束
波前
计算
作者
Feiyi Deng,rui guo,Mohan Qian,Ruirui Zhang,Yunfeng Hu,Yu Lei,Xutao Mo,Junbao Hu
出处
期刊:Applied optics-OT
[Optica Publishing Group]
日期:2026-09-10
卷期号:66 (13): C21-C21
摘要
Imaging through scattering media remains a challenging inverse problem in computational optics. Deep learning enables end to end reconstruction, but existing networks are often too large for portable or embedded systems. We exploited a lightweight network named DSC-U-Net, which replaces standard convolutions in the classic U-Net with depthwise separable convolutions. This decomposition into channel-wise spatial filtering followed by pointwise fusion aligns with the statistical properties of speckle patterns, extracting local textures and cross channel features while dramatically reducing model complexity. Under 600-mesh scattering, DSC-U-Net achieves an average SSIM of 0.900 and PSNR of 25.61 dB, only 1.2% and 4.3% lower than the standard U-Net, yet reduces parameters by 72%, FLOPs by 63%, and model size by 72%. Under 220-mesh, SSIM and PSNR drop by 2.1% and 4.9%. On a CPU, DSC-U-Net runs at 18.7 ms per frame, about 2.5 times faster than U-Net, approaching real time performance. Ablation studies show that computational cost is concentrated in high channel deep layers, and full replacement yields the greatest lightweighting benefit. Our method strikes an excellent balance between reconstruction fidelity and efficiency, offering a practical solution for deploying scattering imaging on resource limited platforms.
科研通智能强力驱动
Strongly Powered by AbleSci AI