Search for: All records

We investigate a convolutional neural network with the first layer implemented by specially designed plasmonic metasurface photodetectors, showing an order of magnitude decrease in computational complexity for the accurate classification of transparent biological cells. 

The development of photonic technologies for machine learning is a promising avenue toward reducing the computational cost of image classification tasks. Here we investigate a convolutional neural network (CNN) where the first layer is replaced by an image sensor array consisting of recently developed angle-sensitive metasurface photodetectors. This array can visualize transparent phase objects directly by recording multiple anisotropic edge-enhanced images, analogous to the feature maps computed by the first convolutional layer of a CNN. The resulting classification performance is evaluated for a realistic task (the identification of transparent cancer cells from seven different lines) through computational-imaging simulations based on the measured angular characteristics of prototype devices. Our results show that this hybrid optoelectronic network can provide accurate classification (>90%) similar to its fully digital baseline CNN but with an order-of-magnitude reduction in the number of calculations. 

Most research on deep learning algorithms for image denoising has focused on signal-independent additive noise. Focused ion beam (FIB) microscopy with direct secondary electron detection has an unusual Neyman Type A (compound Poisson) measurement model, and sample damage poses fundamental challenges in obtaining training data. Model-based estimation is difficult and ineffective because of the nonconvexity of the negative log likelihood. In this paper, we develop deep learning-based denoising methods for FIB micrographs using synthetic training data generated from natural images. To the best of our knowledge, this is the first attempt in the literature to solve this problem with deep learning. Our results show that the proposed methods slightly outperform a total variation-regularized model-based method that requires time-resolved measurements that are not conventionally available. Improvements over methods using conventional measurements and less accurate noise modeling are dramatic - around 10 dB in peak signal-to-noise ratio.