研究目的
To develop a deep neural network capable of simultaneously restoring images destroyed by two distinct scattering media, demonstrating its power in reconstructing images diffused by glass diffuser or multi-mode fiber, and showing generalization ability to reconstruct images not in the training dataset.
研究成果
The deep neural network based on U-net architecture successfully reconstructs images from speckle patterns generated by two distinct scattering media simultaneously, with low MSE and high correlation values. It demonstrates strong generalization ability, reconstructing untrained patterns such as handwritten letters and images from different positions of the diffuser. This approach facilitates optical transmission studies and expands machine learning applications in optics.
研究不足
The study is limited to specific scattering media (glass diffuser and multi-mode fiber) and uses a dataset based on handwritten numbers and letters. The generalization to other types of images or scattering conditions may require additional training data. The experimental setup involves fixed parameters, and variations in media properties could affect performance.
1：Experimental Design and Method Selection:
The study uses a U-net architecture deep neural network for end-to-end image reconstruction from speckle patterns. The network is trained with a blended dataset of speckle-reference image pairs from two scattering media (glass diffuser and multi-mode fiber). The optimizer is Adam with a binary cross-entropy loss function.
2：Sample Selection and Data Sources:
The dataset consists of 2000 handwritten number images from the MNIST database, upsampled to 512x512 pixels. These are used to generate speckle patterns through experimental setups involving scattering media.
3：List of Experimental Equipment and Materials:
Includes a phase-only spatial light modulator (SLM, Hamamatsu, pixel pitch 12.5μm), glass diffuser plate (Thorlabs, 220 Grits), multi-mode optical fiber (Ideaoptics, L=1m, core diameter 600μm, N.A.=0.22), CCD camera, laser (λ=880nm), lenses, polarizers, pinholes, and irises.
4：5μm), glass diffuser plate (Thorlabs, 220 Grits), multi-mode optical fiber (Ideaoptics, L=1m, core diameter 600μm, N.A.=22), CCD camera, laser (λ=880nm), lenses, polarizers, pinholes, and irises.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Images are displayed on the SLM, and the first-order diffraction is passed through scattering media to produce speckle patterns. These patterns are captured by a CCD camera, normalized, and used to train the DNN. The training involves 1000 epochs with a learning rate of 10^-6, using two NVIDIA GTX 1080Ti GPUs and TensorFlow framework.
5：Data Analysis Methods:
Performance is quantified using mean squared error (MSE) and correlation (Corr.) between reconstructed and original images.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
