LeeKosarajuSankaranarayanan - Revision history

imported>Psych2012: /* Conclusion */

2012-03-22T07:02:47Z

Conclusion

← Older revision		Revision as of 07:02, 22 March 2012
Line 264:		Line 264:

	=== Conclusion ===		=== Conclusion ===
	The algorithms and theories explored in this project are proof that near infrared data is useful for post-processing techniques and may allow for automation of computational photography methods. They improve upon the methods currently used, and show that exploration of further uses is worthwhile. ~~It may not, however, be yet~~ the ~~time~~ to ~~rework the imaging pipeline~~ to ~~allow~~ for ~~near infrared data~~ to ~~be captured by~~ digital cameras.		The algorithms and theories explored in this project are proof that near infrared data is useful for post-processing techniques and may allow for automation of computational photography methods. They improve upon the methods currently used, and show that exploration of further uses of NIR in computational photography is worthwhile. Since the setup required to capture NIR is often quite complex, there needs to be a wider gamut of applications than just those discussed here for NIR capture to become a standard feature in digital cameras.

	== Acknowledgements ==		== Acknowledgements ==

imported>Psych2012: /* Explored Ideas */

2012-03-22T06:59:45Z

Explored Ideas

← Older revision		Revision as of 06:59, 22 March 2012
Line 258:		Line 258:

	==== Explored Ideas ====		==== Explored Ideas ====
	Shadow Elimination		Shadow Elimination <br>
	Shadow elimination begins with shadow detection. In order to detect regions of shadows in an image, people have used the basic insight that most dark objects in an image are not dark in all 3 channels (R,G,B) while shadows are mostly dark in all three bands. More robustness can be achieved by using ratios (R/NIR, G/NIR, B/NIR) to indicate presence of shadows. Although simple to implement, we realized that any publication on shadow removal would show examples with no dark objects and a well-lit atmosphere. The images we possessed didn't quite satisfy these requirements - they would either have faces (where hair is dark in all bands) or dark trees (the green would always be detected as shadows by these algorithms). We refer the interested reader to [1].		Shadow elimination begins with shadow detection. In order to detect regions of shadows in an image, people have used the basic insight that most dark objects in an image are not dark in all 3 channels (R,G,B) while shadows are mostly dark in all three bands. More robustness can be achieved by using ratios (R/NIR, G/NIR, B/NIR) to indicate presence of shadows. Although simple to implement, we realized that any publication on shadow removal would show examples with no dark objects and a well-lit atmosphere. The images we possessed didn't quite satisfy these requirements - they would either have faces (where hair is dark in all bands) or dark trees (the green would always be detected as shadows by these algorithms). We refer the interested reader to [1].

imported>Psych2012: /* Explored Ideas */

2012-03-22T06:59:18Z

Explored Ideas

← Older revision		Revision as of 06:59, 22 March 2012
Line 259:		Line 259:
	==== Explored Ideas ====		==== Explored Ideas ====
	Shadow Elimination		Shadow Elimination
	Shadow elimination begins with shadow detection. In order to detect regions of shadows in an image, people have used the basic insight that most dark objects in an image are not dark in all 3 channels (R,G,B) while shadows are mostly dark in all three bands. More robustness can be achieved by using ratios (R/NIR, G/NIR, B/NIR) to indicate presence of shadows. Although simple to implement, we realized that any publication on shadow removal would show examples with no dark objects and a well-lit atmosphere. The images we possessed didn't quite satisfy these requirements - they would either have faces (where hair is dark in all bands) or dark trees (the green would always be detected as shadows by these algorithms). We refer the interested reader to [].		Shadow elimination begins with shadow detection. In order to detect regions of shadows in an image, people have used the basic insight that most dark objects in an image are not dark in all 3 channels (R,G,B) while shadows are mostly dark in all three bands. More robustness can be achieved by using ratios (R/NIR, G/NIR, B/NIR) to indicate presence of shadows. Although simple to implement, we realized that any publication on shadow removal would show examples with no dark objects and a well-lit atmosphere. The images we possessed didn't quite satisfy these requirements - they would either have faces (where hair is dark in all bands) or dark trees (the green would always be detected as shadows by these algorithms). We refer the interested reader to [1].

	Illuminant Detection - The benefits of being able to detect and correct for the illuminant of a scene are well known, and Süsstrunk and Fredembach have suggested methods for accurately approximating the illuminant in a scene. We explored this algorithm, and attempted to recreate and test it, but did not get past the process of estimating the illuminant temperature in the way they describe. The heuristics they use to make the final illuminant decision require knowledge of their data sets or more images in different lighting environments to fully test. That being said, the potential benefits of this are large, but were not fully explored in this process.		Illuminant Detection - The benefits of being able to detect and correct for the illuminant of a scene are well known, and Süsstrunk and Fredembach have suggested methods for accurately approximating the illuminant in a scene. We explored this algorithm, and attempted to recreate and test it, but did not get past the process of estimating the illuminant temperature in the way they describe. The heuristics they use to make the final illuminant decision require knowledge of their data sets or more images in different lighting environments to fully test. That being said, the potential benefits of this are large, but were not fully explored in this process.

imported>Psych2012: /* Future Directions */

2012-03-22T06:58:40Z

Future Directions

← Older revision		Revision as of 06:58, 22 March 2012
Line 255:		Line 255:

	=== Future Directions ===		=== Future Directions ===
	~~The future of this~~ project ~~lies in~~ experimenting with more post-processing algorithms that may benefit from the use of Near Infrared data. If a plethora of techniques can be improved, then it may be cause for a reversal of our conclusion, and a reworking of the imaging pipeline may be worthwhile -- but until that time, research in this field should persist. Some suggested areas, which we as a group managed to begin but not complete are Shadow Elimination and Illuminant Detection.		Future work extending our project could involve experimenting with more post-processing algorithms that may benefit from the use of Near Infrared data. Some suggested areas, which we as a group managed to begin but not complete are Shadow Elimination and Illuminant Detection.

	==== Explored Ideas ====		==== Explored Ideas ====
	Shadow Elimination		Shadow Elimination
			Shadow elimination begins with shadow detection. In order to detect regions of shadows in an image, people have used the basic insight that most dark objects in an image are not dark in all 3 channels (R,G,B) while shadows are mostly dark in all three bands. More robustness can be achieved by using ratios (R/NIR, G/NIR, B/NIR) to indicate presence of shadows. Although simple to implement, we realized that any publication on shadow removal would show examples with no dark objects and a well-lit atmosphere. The images we possessed didn't quite satisfy these requirements - they would either have faces (where hair is dark in all bands) or dark trees (the green would always be detected as shadows by these algorithms). We refer the interested reader to [].

	Illuminant Detection - The benefits of being able to detect and correct for the illuminant of a scene are well known, and Süsstrunk and Fredembach have suggested methods for accurately approximating the illuminant in a scene. We explored this algorithm, and attempted to recreate and test it, but did not get past the process of estimating the illuminant temperature in the way they describe. The heuristics they use to make the final illuminant decision require knowledge of their data sets or more images in different lighting environments to fully test. That being said, the potential benefits of this are large, but were not fully explored in this process.		Illuminant Detection - The benefits of being able to detect and correct for the illuminant of a scene are well known, and Süsstrunk and Fredembach have suggested methods for accurately approximating the illuminant in a scene. We explored this algorithm, and attempted to recreate and test it, but did not get past the process of estimating the illuminant temperature in the way they describe. The heuristics they use to make the final illuminant decision require knowledge of their data sets or more images in different lighting environments to fully test. That being said, the potential benefits of this are large, but were not fully explored in this process.

	~~==== Other Ideas ====~~
	~~More post-processing algorithms~~

	~~Simulating sensor arrays with NIR capturing~~

	=== Conclusion ===		=== Conclusion ===

imported>Psych2012: /* Results */

2012-03-22T06:48:57Z

Results

← Older revision		Revision as of 06:48, 22 March 2012
Line 92:		Line 92:
	* Window size and <math>\sigma</math> for the Bilateral filter.		* Window size and <math>\sigma</math> for the Bilateral filter.
	* <math>\alpha</math> for weighted region mask		* <math>\alpha</math> for weighted region mask
	Our observation was that these parameters exhibit consistent behavior across different images i.e. their value only needs to be changed in a narrow range to get optimal results.<br>		Our observation was that these parameters exhibit consistent behavior across different images i.e. their value only needs to be changed in a narrow range to get optimal results.<br><br>
	<b>Bilateral Filter parameters</b><br>		<b>Bilateral Filter parameters</b><br>
	Since most of the detail is already captured in the Haar difference images, the algorithm produces very similar outputs for a wide range of values for the Bilateral filter parameters.<br>		Since most of the detail is already captured in the Haar difference images, the algorithm produces very similar outputs for a wide range of values for the Bilateral filter parameters.<br><br>
	<b> Run time </b><br>		<b> Run time </b><br>
	The algorithm was found to be very fast and took less than 2 minutes on a machine with Dual-Core AMD Opteron(TM) 2220 processor.<br><br> To summarize this section, we have validated claims that there is better contrast in certain NIR bands than in the visual bands, analyzed an algorithm that exploits this property to enhance photographs using NIR information, presented a way to choose which NIR band to use in case of hyperspectral capture and looked at the performance of the algorithm across different values for the parameters.		The algorithm was found to be very fast and took less than 2 minutes on a machine with Dual-Core AMD Opteron(TM) 2220 processor.<br><br> To summarize this section, we have validated claims that there is better contrast in certain NIR bands than in the visual bands, analyzed an algorithm that exploits this property to enhance photographs using NIR information, presented a way to choose which NIR band to use in case of hyperspectral capture and looked at the performance of the algorithm across different values for the parameters.

imported>Psych2012: /* Results */

2012-03-22T06:47:30Z

Results

← Older revision		Revision as of 06:47, 22 March 2012
Line 96:		Line 96:
	Since most of the detail is already captured in the Haar difference images, the algorithm produces very similar outputs for a wide range of values for the Bilateral filter parameters.<br>		Since most of the detail is already captured in the Haar difference images, the algorithm produces very similar outputs for a wide range of values for the Bilateral filter parameters.<br>
	<b> Run time </b><br>		<b> Run time </b><br>
	The algorithm was found to be very fast and took less than 2 minutes		The algorithm was found to be very fast and took less than 2 minutes on a machine with Dual-Core AMD Opteron(TM) 2220 processor.<br><br> To summarize this section, we have validated claims that there is better contrast in certain NIR bands than in the visual bands, analyzed an algorithm that exploits this property to enhance photographs using NIR information, presented a way to choose which NIR band to use in case of hyperspectral capture and looked at the performance of the algorithm across different values for the parameters.

	== Haze Removal ==		== Haze Removal ==

imported>Psych2012: /* Results */

2012-03-22T06:40:31Z

Results

← Older revision		Revision as of 06:40, 22 March 2012
Line 92:		Line 92:
	* Window size and <math>\sigma</math> for the Bilateral filter.		* Window size and <math>\sigma</math> for the Bilateral filter.
	* <math>\alpha</math> for weighted region mask		* <math>\alpha</math> for weighted region mask
	Our observation was that these parameters exhibit consistent behavior across different images i.e. their value only needs to be changed in a narrow range to get optimal results.		Our observation was that these parameters exhibit consistent behavior across different images i.e. their value only needs to be changed in a narrow range to get optimal results.<br>
	<b>Bilateral Filter parameters</b>		<b>Bilateral Filter parameters</b><br>
	Since most of the detail is already captured in the Haar difference images, the algorithm produces very similar outputs for a wide range of values for the Bilateral filter parameters		Since most of the detail is already captured in the Haar difference images, the algorithm produces very similar outputs for a wide range of values for the Bilateral filter parameters.<br>
	<b> Run time </b>		<b> Run time </b><br>
	The algorithm was found to be very fast and took less than 2 minutes		The algorithm was found to be very fast and took less than 2 minutes

imported>Psych2012: /* Results */

2012-03-22T06:39:28Z

Results

← Older revision		Revision as of 06:39, 22 March 2012
Line 88:		Line 88:
	File:sf_vis_ce.jpg\|Figure 15: Contrast Enhanced Image		File:sf_vis_ce.jpg\|Figure 15: Contrast Enhanced Image
	</gallery>		</gallery>
			We can clearly see from the results that there is good contrast enhancement in the final images. This method also requires very few tweaks, the only parameters required to be set are
			* Choice of NIR band with best contrast
			* Window size and <math>\sigma</math> for the Bilateral filter.
			* <math>\alpha</math> for weighted region mask
			Our observation was that these parameters exhibit consistent behavior across different images i.e. their value only needs to be changed in a narrow range to get optimal results.
			<b>Bilateral Filter parameters</b>
			Since most of the detail is already captured in the Haar difference images, the algorithm produces very similar outputs for a wide range of values for the Bilateral filter parameters
			<b> Run time </b>
			The algorithm was found to be very fast and took less than 2 minutes

	== Haze Removal ==		== Haze Removal ==

imported>Psych2012: /* Particular Implementation */

2012-03-22T06:17:27Z

Particular Implementation

← Older revision		Revision as of 06:17, 22 March 2012
Line 48:		Line 48:
	**''Note: The reference paper does not use ''<math>\alpha</math>'' and this restricts the maximum value for any pixel in ''<math>W</math>'' to be 0.2 which we found to be very low for the images we used. Hence ''<math>\alpha</math>'' was set to a value between 2 and 3 to increase the range of values in ''<math>W</math>		**''Note: The reference paper does not use ''<math>\alpha</math>'' and this restricts the maximum value for any pixel in ''<math>W</math>'' to be 0.2 which we found to be very low for the images we used. Hence ''<math>\alpha</math>'' was set to a value between 2 and 3 to increase the range of values in ''<math>W</math>

	* '''Transfer Contrast''': <br>For reasons mentioned previously, a bilateral filter is applied on both V and N. Let Vb and Nb represent the output of the bilateral filter. The detail images Vd and Nd are calculated as follows<br><center><math> Vd = V - Vb</math><br><math>Nd = N - Nb</math></center><br> A histogram transfer is used to modify Vb based on the histogram of Nb. The values of pixels in Vb are adjusted until the probability distribution function looks like that of the pixels in Nb. Let Vb' be the output of the histogram transfer. A contrast transferred image V' is obtained from the new base layer Vb' using <br><center><math> V' = W.(Vb'+Vd)+(1-W).V </math></center><br> where . is pixel-wise multiplication		* '''Transfer Contrast''': <br>For reasons mentioned previously, a bilateral filter is applied on both ''V'' and ''N''. Let ''Vb'' and ''Nb'' represent the output of the bilateral filter. The detail images ''Vd'' and ''Nd'' are calculated as follows<br><center><math> Vd = V - Vb</math><br><math>Nd = N - Nb</math></center><br> A histogram transfer is used to modify ''Vb'' based on the histogram of ''Nb''. The values of pixels in ''Vb'' are adjusted until the probability distribution function looks like that of the pixels in ''Nb''. Let <i>Vb'</i> be the output of the histogram transfer. A contrast transferred image <i>V'</i> is obtained from the new base layer <i>Vb'</i> using <br><center><math> V' = W.(Vb'+Vd)+(1-W).V </math></center><br> where . is pixel-wise multiplication
	* '''Transfer Texture''': <br>Texture is enhanced from the information in NIR using the following method<br><center><math> VH' = W.NH + (1-W).VH</math></center><br> VV' and VD' are obtained similarly		* '''Transfer Texture''': <br>Texture is enhanced from the information in NIR using the following method<br><center><math> VH' = W.NH + (1-W).VH</math></center><br> <i>VV'</i> and <i>VD'</i> are obtained similarly
	* An inverse Haar transform is applied on <i>V',VH', VV' and VD'</i> to get the enhanced Value image(of HSV). By using the original H and S values we can get an enhanced visual image.		* An inverse Haar transform is applied on <i>V',VH', VV' and VD'</i> to get the enhanced Value image(of HSV). By using the original H and S values we can get an enhanced visual image.

imported>Psych2012: /* Particular Implementation */

2012-03-22T06:15:43Z

Particular Implementation

← Older revision		Revision as of 06:15, 22 March 2012
Line 50:		Line 50:
	* '''Transfer Contrast''': <br>For reasons mentioned previously, a bilateral filter is applied on both V and N. Let Vb and Nb represent the output of the bilateral filter. The detail images Vd and Nd are calculated as follows<br><center><math> Vd = V - Vb</math><br><math>Nd = N - Nb</math></center><br> A histogram transfer is used to modify Vb based on the histogram of Nb. The values of pixels in Vb are adjusted until the probability distribution function looks like that of the pixels in Nb. Let Vb' be the output of the histogram transfer. A contrast transferred image V' is obtained from the new base layer Vb' using <br><center><math> V' = W.(Vb'+Vd)+(1-W).V </math></center><br> where . is pixel-wise multiplication		* '''Transfer Contrast''': <br>For reasons mentioned previously, a bilateral filter is applied on both V and N. Let Vb and Nb represent the output of the bilateral filter. The detail images Vd and Nd are calculated as follows<br><center><math> Vd = V - Vb</math><br><math>Nd = N - Nb</math></center><br> A histogram transfer is used to modify Vb based on the histogram of Nb. The values of pixels in Vb are adjusted until the probability distribution function looks like that of the pixels in Nb. Let Vb' be the output of the histogram transfer. A contrast transferred image V' is obtained from the new base layer Vb' using <br><center><math> V' = W.(Vb'+Vd)+(1-W).V </math></center><br> where . is pixel-wise multiplication
	* '''Transfer Texture''': <br>Texture is enhanced from the information in NIR using the following method<br><center><math> VH' = W.NH + (1-W).VH</math></center><br> VV' and VD' are obtained similarly		* '''Transfer Texture''': <br>Texture is enhanced from the information in NIR using the following method<br><center><math> VH' = W.NH + (1-W).VH</math></center><br> VV' and VD' are obtained similarly
	* An inverse Haar transform is applied on V',VH', VV' and VD' to get the enhanced Value image(of HSV). By using the original H and S values we can get an enhanced visual image.		* An inverse Haar transform is applied on <i>V',VH', VV' and VD'</i> to get the enhanced Value image(of HSV). By using the original H and S values we can get an enhanced visual image.

	==== NIR Band Selection ====		==== NIR Band Selection ====