Fig. 1. Physical formation of blurring artifacts under conventional and coded exposure settings, and analysis in spatial and frequency domains.

Fig. 2. Overall flowchart of the proposed framework. The coded exposure imaging system and the learning-based deblurring algorithm are respectively modeled with an optical blur encoder and a computational blur decoder, and together form an end-to-end differentiable forward model. In the training stage, the parameters of the whole model are optimized together through gradient descent until convergence. In the inference stage, the learned encoding sequence will be loaded to the controller of the camera shutter (or its equivalent), and the computational blur decoder will be employed to deblur the captured coded blurry images.

Fig. 3. Architecture of the deblurring neural network DeepRFT [29] in the proposed framework.

Fig. 4. Prototype system for coded exposure photography. It employs a liquid crystal element to serve as an external shutter for exposure encoding.

Fig. 5. Synthesized blurry images under different exposure encoding settings and corresponding deblurring results. (Please zoom in for a better view.)

Fig. 6. Frequency spectra of different encoding sequences.

Fig. 7. Influence of the encoding sequence’s length on the deblurring performance of the proposed framework.

Fig. 8. Real-captured blurry images under exposure with different encoding sequences and corresponding deblurring results. (Please zoom in for a better view.)