Adaptive Image and Video Retargeting Technique Based on Fourier Analysis

Jun-Seong Kim, Jin-Hwan Kim, and Chang-Su KimSchool of Electrical Engineering, Korea University, Seoul, Korea

Reporter: Chia-Hao HsiehDate: 20100316

2

Outline

• Introduction• Methods• Simulation results

3

Introduction

Original Scaled Cropped Retargeted

4

Introduction

5

Methods

• Partitioning• Scaling distortions• Adaptive scaling

6

Partitioning

K = 10 in this work

7

Scaling distortions

• z[n]: a row signal of length lk in the kth strip

• Reduce the length by removing rk pixels to obtain a downsampled signal zd[n]

• The sampling rate is reduced by a factor of (1-(rk/lk))π

8

Scaling distortions

• The signal should be lowpass-filtered with a cutoff frequency ωc = (1-(rk/lk))π to avoid aliasing artifacts

• Scaling distortions (The energy of the lost high frequency components)

Can be approximated by

Except for the dominant DC component

A smaller sk makes the exponential function decrease more quickly

9

Scaling distortions

Can be approximated by

Scaling distortion of the kth strip:

Except for the dominant DC component

10

Adaptive scaling

To minimize the sum of the distortions of strips

R = Ws − Wt

Constrained optimization problemCan be solved by minimizing the Lagrangian cost function

11

Adaptive scaling

12

Extension to video retargeting

• Cut the volume into parallelepipeds along the time axis

• Each parallelpiped is scaled down spatially Partitioning of a video sequence

for horizontal resizing.

13

Simulation Results

14

Simulation Results

15

Conclusions

• The proposed algorithm requires much less computations than the seam carving

• Excellent temporal coherence without jitter artifacts