BM3D Denoised (b) (c) Reconstruction BM3D Denoised (e) (f) Prototype 1 Reconstruction Prototype 2 (a) 3 mm (d) Figure 6. A FlatCam prototype comparison. (a) Our first prototype with the chrome mask placed directly in front of the Flea3 sensor; (d) is our second prototype with the Omnivision sensor directly epoxied to mask (insets show close-up of the sensors and masks). (b) and (e) are reconstructions and (c) and (f) are BM3D denoised reconstructions for first and second prototype, respectively. The smaller feature size and pixel pitch of the Omnivision prototype provide superior resolution at the cost of more noisier image reconstruction. Reconstruction Performance for a Programmable Mask 16 One Acquisition PSNR = 8.66 dB Nine Acquisitions PSNR = 13.67 dB 36 Acquisitions PSNR = 15.75 dB PSNR = 8.66 dB PSNR = 11.52 dB PSNR = 10.50 dB Static 15 13 12 11 Static 0.05 px/Acquisition 0.1 px /Acquisition 0.2 px/Acquisition Constant Mask 10 9 8 1 4 8 12 16 20 0.2 px/Frame PSNR (dB) 14 24 28 32 36 Number of Image Acquisitions (Masks) (a) (b) Figure 7. A simulation experiment of a FlatCam with a programmable mask. (a) Reconstruction performance (in terms of peak-signal-to-noise ratio, PSNR) as we increase the number of image acquisitions (masks). A different mask pattern is used for each acqusition. The PSNR increases consistently for static and slow-moving scenes, but after peaking early, deteriorates for faster moving scenes due to model mismatch. The green dot indicates the performance when all nine image acquisitions of a static scene are done using the same (constant) mask. Though the performance is better than a single acquisition, it is still outshined by the programmable mask. (b) Reconstructed images using one, nine, and 36 acquisitions for the static scene and fast-moving scene. 32 IEEE SIgnal ProcESSIng MagazInE | September 2016 |