13
Computational Photography Final project – Deblur 601415004 巫巫巫

Computational Photography Final project – D eblur

  • Upload
    karlyn

  • View
    60

  • Download
    0

Embed Size (px)

DESCRIPTION

Computational Photography Final project – D eblur. 601415004 巫承熹. Background. In photography, it have different type of blur Camera shake( 相機晃動 ) User moving hands Scene motion( 場景位移 ) Objects in the scene moving Defocus blur ( 失焦 ) Depth of field effect. - PowerPoint PPT Presentation

Citation preview

Page 1: Computational Photography  Final project –  D eblur

Computational Photography Final project – Deblur601415004巫承熹

Page 2: Computational Photography  Final project –  D eblur

BackgroundIn photography, it have different type of blur

Camera shake(相機晃動 )• User moving hands

Scene motion(場景位移 )• Objects in the scene moving

Defocus blur (失焦 )• Depth of field effect

Prof. Shing-Min Liu, Computational Photography: Applied Graphics and Imaging course lecture(2013) 

Page 3: Computational Photography  Final project –  D eblur

Introduction

在拍照攝影時,有上述幾種造成影像模糊的原因。其中,相機的晃動,可以藉由裝設感應器(如:加速度計、陀螺儀 )來擷取移動向量,以解決模糊。這類的去模糊方法已經廣泛的被應用在相機的光學防手震 (OIS)上。

HTC官方網站 , http://www.htc.com/tw/zoe/stabilization/

Page 4: Computational Photography  Final project –  D eblur

Introduction

然而,場景 (物件 )的位移卻無法直接由相機上的感應器直接的量測,因此,在這裡我們試著用基本原理,去除場景 (物件 )的簡單位移造成的模糊。

Page 5: Computational Photography  Final project –  D eblur

Estimation the Degradation Function Give the blur source image:

We try to recover the image using estimation the degradation funtiong(x,y) = f(x,y) * h(x,y) (1)G(u,v) =F(u,v) * H(u,v) (2)

Page 6: Computational Photography  Final project –  D eblur

MethodStep1: Load the blur source image , then convert it to the

frequency domain using FFT. Get .Step2: Calculate , in frequency domain. In this case, we only have horizontal motion, therefore, , which , and a is parameter between 0.11 to 0.15.Step3: for avoiding case of “division zero”, we set a threshold , which is a small positive constant. If , then let ,

otherwise, .Step4: Use inverse FFT to convert from frequency domain to spatial domain .

Page 7: Computational Photography  Final project –  D eblur

FFT

Flow Cart Load Image File

𝐹 (𝑢 ,𝑣 )= 𝐺 (𝑢 ,𝑣 )𝐻 (𝑢 ,𝑣)

Inverse FFT

Display

Page 8: Computational Photography  Final project –  D eblur

User Interface ( Qt )

QT是一個跨平台的 C++應用程式開發框架,被廣泛的應用於開發 GUI程式。選擇使用這個 GUI的介面,除了他擁有影像處理能力,還有因為它的支援跨平台的特性。這使我們可以將程式移植至嵌入式平台上執行,以提升系統攜帶的方便性與實用性。

Page 9: Computational Photography  Final project –  D eblur

User InterfaceLoad Image file (.bmp only)

Source Image Display here Result Image Display

here

Click the button to display reult

Set the parameter

Page 10: Computational Photography  Final project –  D eblur

Result

因為每個來源影像的 blur kernel 是未知的,因此根據不同的來源影像,會有不同的最佳輸入參數。

a=0.14 a=0.11

Page 11: Computational Photography  Final project –  D eblur

Result

然而,並不是每個 case都可以成功地找到參數,使圖片去模糊。在這個 case中,圖片經過轉換產生的許多雜訊。但雖然如此,但是車子的下半部分仍然成功地去模糊。

a=0.156

Demo Video( Youtube, Demo.avi )

Page 12: Computational Photography  Final project –  D eblur

結論在照相越來流行的時代, Deblur是非常實用的領域,目前市面上利用OIS解決相機晃動的作法已是非常普及,然而拍攝物體的移動是未知的,因此,我們運用預測 Degradation Function的方式來去除模糊。目前我們以手動調整參數來實現。未來,可以利用一些設定參數的策略,來完善這個研究。

Page 13: Computational Photography  Final project –  D eblur

Reference

[1] Prof. Shing-Min Liu, Computational Photography: Applied Graphics and Imaging course lecture(2013) 

[2] HTC官方網站 , http://www.htc.com/tw/zoe/stabilization/

[3] Wikipedia, http://zh.wikipedia.org/zh-hant/Qt

[4] M. Ben-Ezra and S. K. Nayar  "Motion-based motion deblurring",  IEEE Trans. Pattern Anal. Mach. Intell.,  vol. 26,  no. 6,  pp.689 -698 2004