Deblurring sensor images using known psf models - Revision history

imported>Projects221: /* Appendix II */

2014-03-21T06:27:24Z

Appendix II

← Older revision		Revision as of 06:27, 21 March 2014
Line 216:		Line 216:
	== Appendix II ==		== Appendix II ==

	Our group met frequently to discuss and understand ISET before setting up the simulation infrastructure for the experiments. Then we carried out the experiments in parallel and analysed each part individually and as a group to draw concrete conclusions. We divided the project workload very broadly as listed below:		Our group met frequently to discuss and understand ISET (mainly Rakesh Ramesh) before setting up the simulation infrastructure for the experiments. Then we carried out the experiments in parallel and analysed each part individually and as a group to draw concrete conclusions. We divided the project workload very broadly as listed below:

	Rakesh Ramesh: Shift-invariant PSF study, Wavelength-dependent PSF study		Rakesh Ramesh: Shift-invariant PSF study, Wavelength-dependent PSF study

imported>Projects221: /* Conclusions */

2014-03-21T06:17:53Z

Conclusions

← Older revision		Revision as of 06:17, 21 March 2014
Line 192:		Line 192:
	== Conclusions ==		== Conclusions ==

	The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF ~~as a function of the~~ PSF ~~model~~ and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give considerable further improvement, mainly because the edge artifacts start to have greater effect on the quality. In the wavelength dependent ~~case and in the color image~~ we didn't manage to get a deblurred image that was improved compared to the sensor image, something that is attributed to the effects of the bayer filter.		The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF under different PSF models and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give considerable further improvement, mainly because the edge artifacts start to have greater effect on the quality. In the wavelength dependent PSF study we didn't manage to get a deblurred image that was improved compared to the sensor image, something that is attributed to the effects of the bayer filter. The noise margin analysis also provides us the 3 interesting inferences: the sensor image can tolerate read noise to an extent for deblurring, the sensor image and deblurring is not affected by the PRNU level and finally the sensor image and deblurring is heavily dependent on the DSNU level which plays a big factor in quality of the image.

			To conclude, image restoration is a rapidly growing field that has helped correct the physical limitations in the modern lens technology, one of which is blurring. In this project, we have unconvered some of the aspects to deblurring the image produced by the sensor and under different PSF models and noise margins and we believe that there is still a lot to explore in this area to get very precise accurate images that are closer to the original scenes for a better image system for the future.

	== References ==		== References ==

imported>Projects221: /* References */

2014-03-21T06:12:54Z

References

← Older revision		Revision as of 06:12, 21 March 2014
Line 197:		Line 197:

	* Wikipedia : Point Spread Function		* Wikipedia : Point Spread Function
			* Wikipedia : Peak Signal to Noise Ratio (PSNR)
			* Wikipedia : Structural Similarity Index (SSIM)
			* Mathworks : Lucy-Richardson Algorithm (deconvlucy)
	* Image Enhancement using Calibrated Lens Simulations (Y.Shih,B.Guenter and N.Joshi)		* Image Enhancement using Calibrated Lens Simulations (Y.Shih,B.Guenter and N.Joshi)
	* Image Capture Simulation Using an Accurate and Realistic Lens Model (R.Short,D.Williams and A.Jones)		* Image Capture Simulation Using an Accurate and Realistic Lens Model (R.Short,D.Williams and A.Jones)

imported>Projects221: /* Noise Margins */

2014-03-21T06:09:11Z

Noise Margins

← Older revision		Revision as of 06:09, 21 March 2014
Line 171:		Line 171:

	=== Noise Margins ===		=== Noise Margins ===
	The deconvolution algorithm used to deblur the sensor image is sensitive to noise. Non-ideal sensor adds noise to the sensor image. ~~Types~~ of ~~Noise~~:		The deconvolution algorithm used to deblur the sensor image is sensitive to noise. Non-ideal sensor adds noise to the sensor image. The different types of noise that affects the sensor are given below:

	* Shot Noise		* Shot Noise
Line 178:		Line 178:
	* Dark Signal Non-Uniformity (DSNU)		* Dark Signal Non-Uniformity (DSNU)

	Noise Margin analysis characterizes the sensitivity of the deblurring scheme to the various types of noise. We ~~carried out~~ experiments ~~with different~~ magnitudes of Read, PRNU and DSNU noise. In the ~~following plots we observe~~ that the ~~PSNR metric responds~~ in ~~the same way~~ for ~~both~~ the sensor and deblurred ~~images, i.e. there~~ is ~~a strong correlation between the two as expected.~~		Noise Margin analysis characterizes the sensitivity of the deblurring scheme to the various types of noise. We extend the simulation infrastructure for our previous experiments and vary the magnitudes for the Read, PRNU and DSNU noise and observe the quality of our sensor and deblurred image. We do not vary the shot noise since that is a characteristic that is dependent on the number of photons that are present and is modeled in our noise margin analysis by default. The plots for the PSNR of sensor and the deblurred image vs the noise levels and different types is shown below:

	<center>		<center>
Line 187:		Line 187:
	</gallery>		</gallery>
	</center>		</center>
			<br>
	Note that similar trends can be ~~seen~~ for the shift-variant case.		There are a number of interesting observations we derive from our plots. We observe that the PSNR metric responds in the same way for both the sensor and deblurred images, i.e. there is a strong correlation between the two as expected. Secondly, we see that the sensor image degrades much faster with the read noise while our deblurred is able to withstand the read noise to a certain degree until the image completely degrades (100x). However, we see no noticeable pattern for the PSNR of the sensor and the deblurred image when we varied the PRNU level which implies that the non-uniformity is not a big factor when we consider deblurring. Finally, the DSNU noise in the sensor is the biggest quality determining factor. The quality of the sensor and deblurred image decreases sharply with increase in DSNU noise. Note that similar trends can be observed for the SSIM metric and the shift-variant PSF case.

	== Conclusions ==		== Conclusions ==

imported>Projects221: /* Conclusions */

2014-03-21T05:54:17Z

Conclusions

← Older revision		Revision as of 05:54, 21 March 2014
Line 192:		Line 192:
	== Conclusions ==		== Conclusions ==

	The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF as a function of the PSF model and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give considerable further improvement, mainly because the edge artifacts start to have greater effect on the quality. In the wavelength dependent case and in the color image we didn't manage to get a deblurred image that was improved compared to the sensor image, something that is attributed to the effects of the bayer ~~sensor~~.		The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF as a function of the PSF model and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give considerable further improvement, mainly because the edge artifacts start to have greater effect on the quality. In the wavelength dependent case and in the color image we didn't manage to get a deblurred image that was improved compared to the sensor image, something that is attributed to the effects of the bayer filter.

	== References ==		== References ==

imported>Projects221: /* Conclusions */

2014-03-21T05:53:59Z

Conclusions

← Older revision		Revision as of 05:53, 21 March 2014
Line 192:		Line 192:
	== Conclusions ==		== Conclusions ==

	The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF as a function of the PSF model and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give considerable further improvement, mainly because the edge artifacts start to have greater effect on the quality.		The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF as a function of the PSF model and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give considerable further improvement, mainly because the edge artifacts start to have greater effect on the quality. In the wavelength dependent case and in the color image we didn't manage to get a deblurred image that was improved compared to the sensor image, something that is attributed to the effects of the bayer sensor.

	== References ==		== References ==

imported>Projects221: /* Conclusions */

2014-03-21T05:48:55Z

Conclusions

← Older revision		Revision as of 05:48, 21 March 2014
Line 192:		Line 192:
	== Conclusions ==		== Conclusions ==

	The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF as a function of the PSF model and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give further improvement, mainly because the edge artifacts start to have greater effect on the quality.		The scope of the project was to provide answers on how well we can sharpen an image if we know the underlying PSF as a function of the PSF model and the different noise parameters. Given a field position independent PSF and no additional sensor noise we observed that we could achieve a significant improvement in the sharpness of the image. In the case of shift-variant PSFs we had to partition the image and deblur by using a different PSF for each part. A small partitioning has a significant impact on the PSNR but partitioning the image into very small images does not give considerable further improvement, mainly because the edge artifacts start to have greater effect on the quality.

	== References ==		== References ==

imported>Projects221: /* Appendix I */

2014-03-21T05:45:01Z

Appendix I

← Older revision		Revision as of 05:45, 21 March 2014
Line 207:		Line 207:


	We also had to modify some o the already ~~write~~ ISET scripts. More specifically, the lines 119-121 of the \opticalimage\optics\shiftinvariant\siSynthetic.m have to be commended.		We also had to modify some of the already written ISET scripts. More specifically, the lines 119-121 of the \opticalimage\optics\shiftinvariant\siSynthetic.m have to be commended.

	== Appendix II ==		== Appendix II ==

imported>Projects221: /* PSF Model: Wavelength-dependent; Scene: Lenna */

2014-03-21T05:41:51Z

PSF Model: Wavelength-dependent; Scene: Lenna

← Older revision		Revision as of 05:41, 21 March 2014
Line 158:		Line 158:

	=== PSF Model: Wavelength-dependent; Scene: Lenna ===		=== PSF Model: Wavelength-dependent; Scene: Lenna ===
			We extend the wavelength-dependent study for a larger and denser chromatic image of Lenna. The scene used in our experiments consists of only 3 wavelengths (R-650, G-550, B-450) of the original image to speed-up the blurring and deblurring process. The original scene, the sensor and the deblurred images are as shown below:

	<center>		<center>
Line 166:		Line 167:
	</gallery>		</gallery>
	</center>		</center>
			<br>
			We notice that the sensor image is significantly blurred first by the optics and then by the bayer filter. The bayer filter destroys the structural similarity to the original scene as seen with our SSIM and PSNR metrics. The deblurred image isn't any different from the sensor image and the effects of the bayer filter could not be undone to produce better deblurring.

	=== Noise Margins ===		=== Noise Margins ===

imported>Projects221: /* Noise Margins */

2014-03-21T05:39:56Z

Noise Margins

← Older revision		Revision as of 05:39, 21 March 2014
Line 175:		Line 175:
	* Dark Signal Non-Uniformity (DSNU)		* Dark Signal Non-Uniformity (DSNU)

	Noise Margin analysis characterizes the sensitivity of the deblurring scheme to the various ~~type~~ of noise.		Noise Margin analysis characterizes the sensitivity of the deblurring scheme to the various types of noise. We carried out experiments with different magnitudes of Read, PRNU and DSNU noise. In the following plots we observe that the PSNR metric responds in the same way for both the sensor and deblurred images, i.e. there is a strong correlation between the two as expected.

	<center>		<center>