This talk reviews approaches to combine compressive imaging systems and neural network algorithms for hyperspectral machine vision tasks and how this can be implemented in reconstruction as well as directly on compressive measurements. The first example discussed will be where the spatial light modulator performs the optical computation in the first layer of a convolutional neural network. Next, the same optical system has been used for object recognition utilizing a neural network as a nonlinear transform between the binary patterns on the modulator and the grayscale patterns learned from secants in manifold space. The result is that this task can be done both with fewer patterns and be made much more robust to noise. A new dynamic sampling rate approach to training neural networks specifically for an arbitrary number of compressive measurements in these optical will also be demonstrated. Lastly, our new L1 compressive foveation result will be presented and shown how it can be used to parallelize reconstruction of large images as well as showing that the optimization algorithm can be replaced with a neural network for very fast reconstruction.
Back to Workshop IV: Multi-Modal Imaging with Deep Learning and Modeling