0.3 Summary of approach

Summary of approach

Our approach includes three parts: 1. generation of blurred images, 2. deblur, 3. evaluation.

To generate the blurred images, we first convolve the impulse response of the boxcar and coded filter with the original image. Then, we addzero mean Gaussian noise to the blurred images to simulate the thermal and readout noise caused by real-world camera sensors.

We implemented the deblurring process for both the traditional shutter (boxcar filter) and the flutter shutter (coded filter) in both time andfrequency domains. In time domain implementation, we use least square estimation to solve the linear equation $B=AX+n$ for $X$ . In the above equation, $B$ is the blurred image, $A$ is the smear matrix corresponding to the filter type, and n is noise. In the frequencydom $$ ain implementation, we divide the $\text{DFT}$ of the blurred image by the $\text{DFT}$ of the filter response, and take the inverse $\text{DFT}$ to get our deblurred image.

After getting the deblurred images, we calculated the $\text{PSNR}$ (explained in Results Section) to characterize how well each approach works. We plotted the $\text{PSNR}$ with respect to the variance of the Gaussian noise we added to the blurred image to determine thedeblurring performance of the traditional shutter (boxcar filter) and the flutter shutter (coded filter).