CodedRecon: Video reconstruction for coded exposure imaging techniques

Abstract We present CodedRecon, a deep learning based framework for video reconstruction from coded exposure imaging techniques. It’s a fully differentiable framework consisting of a coded exposure sensor simulation module and a deep-neural network module that learns to reconstruct the video sequence from the input coded exposure measurement. The reconstruction neural network is fully-convolutional and incorporates a spatially varying convolutional layer for exposure aware feature extraction from coded exposure measurements. The users can input measurements from both global and pixel-wise coded exposure techniques and reconstruct a video sequence with 16 frames per input measurement. The framework can be used for benchmarking various coded exposure techniques and the framework’s spatially varying convolutional layer can facilitate further research in the field.