Reconstructing lensless image with ML models and deploying them onto embedded systems
Lensless imaging provides a large panel of benefits : cost, size, weight, etc., that are crucial for wearable application, IoT or medical devices. Such setups require the design of reconstruction algorithms to recover the image from the captured measurements. Most of the current SoTA recon- struction models use deep learning, but the results provided are hardly reproducible and mostly not meant to be deployed into embedded systems. In this work, we implement the work of Monakhova et al. that use uniquely deep learning, and the work of by Khan et al. We then present a way to transform these models to be deployable using TensorFlow Lite, and evaluate the benefits of model optimization techniques such as quantization-aware training(QAT), weight pruning, or weight clustering.
master_thesis_jreymond.pdf
preprint
openaccess
copyright
10.4 MB
Adobe PDF
82c4fe66260c86677e99c3d1226405ad