ECE 438 Introduction

1. ECE 438 Introduction Jan P. Allebach 8 January 2018

2. What is DSP? • Analog signals • Discrete-Time Signals • Digital Signals • Application Scenarios • Process an input signal to generate an output signal • Process an input signal to make a decision

3. What are the applications of DSP?

4. Why is DSP important? • Advantage of processing binary signals • Reproducibility • Bistable devices • Analog signal processing vs. digital signal processing • Analog components are generally more bulky and expensive • DSP is inherently more flexible and more reconfigurable • Discrete mathematical operations vs. analog circuit design • DPS has piggy-backed on the evolution of computer technology • Moore’s Law • Internet

5. OK, but is this stuff really useful?(1/2) I was a student of yours (gasp!) almost 20 years ago now…  Very fond memories of EE 438 and my time at Purdue.                 I’m now working at a military contractor in Cincinnati.  I stayed in Digital Communications – working on Spacecraft Communications links now. We’ve built the docking link for vehicles visiting the Int’l Space Station, as well as a variety of datalinks for Mars and Low Earth Orbit satellites.  Very thankful for the places that my Purdue degree has allowed me to go! We have a job opening in Image Processing, for our Infrared Camera Product line - and I was wondering if you know of any current/former students who might be interested?  . Ken Fischer, 6 December 2017

6. OK, but is this stuff really useful? (2/2) I want to update you with my recent status at Facebook. It’s been 1.5 year since I graduated from Purdue and joined Facebook. Last August I was promoted to Senior Research Scientist (E5) and I’m now tech leading a small team. At Facebook I spent most of my time working on building large-scale ads recommendation algorithm using engagement activities on various types of media on Facebook. On a daily basis we extract high-dimension signals from millions of ad images and videos on Facebook and encode them in deep learner which then predicts probabilities of the users engaging with the candidate ads. Recently we launched a product Value Optimization which helps eCommerce advertisers find users on Facebook who would like to spend more money on their products. It is now generating millions of dollars for Facebook every day. With that said, I found the problems I’m dealing with day-to-day are not so much different from my graduate work and I still find the methodology and problem-solving skills learned at Purdue very useful. Therefore I don’t want to limit my vision to only the problem I’m solving at work. I want to keep close connection with the application driven studies in academia like problems we solved at EISL. Yucheng Liu, 5 January 2018

7. ECE 438 Course Logistics • Course information document • Course syllabus • Course website

8. ECE 438 Staff • Ikbeom Jang – Lab TA • Wanling Jiang – Course TA • Daulet Kenzhebalin – Lab TA Daulet Kenzhebalin Ikbeom Jang Wanling Jiang

9. Enrollment in ECE 438 lab sections Tuesday - 8:30-11:20: 11 students (Sec. 2) Tuesday - 11:30-2:20p: 15 students (Sec. 5) Thursday - 11:30-2:20p: 15 students (Sec. 3) Friday – 11:30-2:20p: 11 students (Sec. 6)

10. Applications of Fourier analysis to halftoning* Halftoning: The process of rendering a continuous tone image with limited number of output levels Three important types of halftones Continuous- tone image Dispersed-dot Periodic clustered-dot Stochastic clustered-dot * Madhur Gupta, AT&T Services, San Ramon, CA Smooth appearance Stable (Fewer printer-induced artifacts) Visible periodic moiré Stable (Fewer printer-induced artifacts) No periodic moiré Good detail rendition Less stable

11. Frequency response of an inkjet printer* Modulation frequency fM Spatial Projection RGB to XYZ 2D XYZ 2D RGB Scanned test target page 1D XYZ 1Dmodulation signal FFT spectrum XYZ to L*a*b* Color Space Projection Extract magnitude at fM FFT 1D L*a*b* * Young Jang, Qualcomm, San Diego, CA

12. Sample Data at Three Frequencies Scanned patches L*a*b* space Modulation signals FFT Spectra EAP (L*a*b*) Mag (peak-peak) f = 20cpi f = 50cpi EAP (L*a*b*) Mag (peak-peak) EAP (L*a*b*) Mag (peak-peak) f = 80cpi Freq(cycles/inch) Distance(inch)

13. Resulting MTF measurement • 1st row of the WM page. 10 20 30 40 50 60 80 100 120 MTF Freq(cycles/inch)

14. Texture analysis using steerable filter banks* • Some Areas in the image which exhibit major orientation texture mask the visibility of banding in different ways • Retrieve the object with uniform texture orientation from the image Horizontal oriented texture: Disrupt the continuity of banding Complex texture: * Yucheng Liu, Purdue ECE graduate student

15. Motivation • Some Areas in the image which exhibit major orientation texture mask the visibility of banding in different ways • Retrieve the object with uniform texture orientation from the image Vertical oriented texture: Conceal the banding in its vertical structure

16. Building Block: Steerable Filter Pyramid • Make a radial and angular subband decomposition of the image in the 2-D frequency domain • The response of the image lightness in each subband will be used in the orientation segmentation

17. Realization of Steerable Filter The Design and Use of Steerable Filters, William T. Freeman and Edward H. Adelson (1991) By Interpolation (Upsampling) By Interpolation of the filter coefficient, we obtain subband decomposition in different radial frequency bands

18. Subband Decomposition Example(Orientation Decomposition)

19. Steerable Pyramid decomposition of Real Image

20. Subband Decomposition Example(Scale Decomposition)

21. Subband Decomposition Example(Scale Decomposition)