s ingle -i mage r efocusing and d efocusing wei zhang, nember, ieee, and wai-kuen cham, senior...

23
SINGLE-IMAGE REFOCUSING AND DEFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

Upload: gannon-marker

Post on 14-Dec-2015

216 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

SINGLE-IMAGE REFOCUSING AND DEFOCUSINGWei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

Page 2: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

OUTLINE

Introdution Background And Problem Fomulation Edge-Based Focus-Map Estimation Image Refocusing By Blind Deconvolution Experiment And Discussions Conclusion

Page 3: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

OUTLINE

Introdution Background And Problem Fomulation Edge-Based Focus-Map Estimation Image Refocusing By Blind Deconvolution Experiment And Discussions Conclusion

Page 4: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

INTRODUCTION

One is image refocusing, which is to recover the sharpness of the blurry defocused objects in an input image and to generate a virtual all-focused image.

The other is defocusing, which is to blur an image and to create defocus effects.

The major novelty of this paper is introducing a tractable SBD framework, which can apply separately these two constraints to ensure image sharpness and to suppress ring artifacts.

Page 5: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

INTRODUCTION

(a) Input narrow-aperture image focusing on the foreground object. (b) Synthesized image with shallower DOF. (c) Synthesized all-focused image. (d) Synthesized image focusing on the background. (e) Detected focus mask (white: defocused regions, black: focused

regions, gray: focus boundaries). (f) Close-up comparison. Left: removing the lens blur using the lens

deblurring in smart sharpen of Photoshop. Right: Our refocused result.

Page 6: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

OUTLINE

Introdution Background And Problem Fomulation Edge-Based Focus-Map Estimation Image Refocusing By Blind Deconvolution Experiment And Discussions Conclusion

Page 7: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

BACKGROUND AND PROBLEM FORMULATION

Image Model

Such blurring process is often modeled as the convolution of the focused image F with PSF.

I denotes the defocused image and n is the noise term

Page 8: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

BACKGROUND AND PROBLEM FORMULATION

Due to the diffraction and aberration of the camera lens, the PSF is approximated normally by a 2-D Gaussian filter, given by

The spread parameter , which is related to the distance of the object to the focal plane, determines the blurriness of the captured image.

Page 9: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

BACKGROUND AND PROBLEM FORMULATION

Edge Modeling A step edge x0 at can be represented by e(x;b,c,x0) = cU(x - x0) + b, where U(‧) is the unit step function. b denotes the edge basis. c represents the edge contrast.

Page 10: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

BACKGROUND AND PROBLEM FORMULATION

The typical edge s(x;b,c,w,x0) can be regarded as a smoothed step edge e(x;b,c,x0), which is obtained by convolving with the 1-D Gaussian filter

and therefore is as follows

Parameters

Page 11: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

OUTLINE

Introdution Background And Problem Fomulation Edge-Based Focus-Map Estimation Image Refocusing By Blind Deconvolution Experiment And Discussions Conclusion

Page 12: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

EDGE-BASED FOCUS-MAP ESTIMATION

Based on the above edge model, a method is proposed in this section to estimate automatically the focus map for an image containing a mixture of focused and defocused objects.

Our proposed method is also based on this assumption and thus shares the common limitation that it cannot estimate an accurate focus map for natural blurry objects such as clouds and shadows.

By contrast, our proposed method is simpler and has lower computational complexity since all edge parameters are derived in closed form.

Page 13: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

EDGE-BASED FOCUS-MAP ESTIMATION

Page 14: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

OUTLINE

Introdution Background And Problem Fomulation Edge-Based Focus-Map Estimation Image Refocusing By Blind Deconvolution Experiment And Discussions Conclusion

Page 15: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

IMAGE REFOCUS BY BLIND DECONVOLUTIOB

1) It is utilized for PSF estimation.

2) An edge sharpness prior is developed to constrain the PSF not to blur the edges and enforce the refocusing image to agree with the precalculated sharpened image in the vicinity of edges.

3) The proposed SBD method will be presented by assuming that the PSF is spatially invariant for the sake of simplicity.

Page 16: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

IMAGE REFOCUS BY BLIND DECONVOLUTIOB

Let be F the refocused image of the blurry image I . F Is expected to satisfy two conditions:

(1) The edges should become sharpened in F (2) The locally smooth regions in I should remain almost unchanged in F .

Page 17: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

OUTLINE

Introdution Background And Problem Fomulation Edge-Based Focus-Map Estimation Image Refocusing By Blind Deconvolution Experiment And Discussions Conclusion

Page 18: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

EXPERIMENTS AND DISCUSSIONS

(a) Estimated focus map. (b) Blurriness of the pixels at the dashed line of (a). (c) Segmentation result based on (a).

Page 19: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

EXPERIMENTS AND DISCUSSIONS

Page 20: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

EXPERIMENTS AND DISCUSSIONS

Page 21: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

OUTLINE

Introdution Background And Problem Fomulation Edge-Based Focus-Map Estimation Image Refocusing By Blind Deconvolution Experiment And Discussions Conclusion

Page 22: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

CONCLUSION

The proposed SBD is free of user initialization and has low computational complexity.

A wide variety of images has been tested to validate the proposed algorithm.

In the future, we intend to use optimized GPU programming to accelerate further the speed of the algorithm.

Finally, we would like to extend the basic idea of this paper to solve other low-level vision problems such as spatially variant deblurring.

Page 23: S INGLE -I MAGE R EFOCUSING AND D EFOCUSING Wei Zhang, Nember, IEEE, and Wai-Kuen Cham, Senior Member, IEEE

THANK YOU FOR YOUR LISTENING