2. Only for
Maxus
Содержание
Введение
Простые модели
Adaptive Fuzzy Post-Filtering
DCT Re-application
Классификация
Support Vector Regression
Modified Mean-Removed Classified Vector Quantization
Регуляризация
Заключение
2
CS MSU Graphics & Media Lab (Video Group)
3. Only for
Maxus
Введение
3
CS MSU Graphics & Media Lab (Video Group)
4. Only for
Maxus
Введение
4
CS MSU Graphics & Media Lab (Video Group)
5. Only for
Maxus
Содержание
Введение
Простые модели
Adaptive Fuzzy Post-Filtering
DCT Re-application
Классификация
Support Vector Regression
Modified Mean-Removed Classified Vector Quantization
Регуляризация
Заключение
6
CS MSU Graphics & Media Lab (Video Group)
6. Only for
Maxus
Adaptive Fuzzy Post-Filtering
j x j (x c , x j )
N
~
xc 1
j (x c , x j )
N
1
/ 2 2
(a , b ) e (a b )
2
2 сила фильтра,
зависит от типа блока
~
x c новое значение пикселя
Adaptive Fuzzy Post-Filtering for Highly Compressed Video. Hao-Song Kong,
Yao Nie, Anthony Vetro, Huifang Sun, Kenneth E. Barner. ICIP 2004.
7
CS MSU Graphics & Media Lab (Video Group)
7. Only for
Maxus
Adaptive Fuzzy Post-Filtering
Adaptive Fuzzy Post-Filtering for Highly Compressed Video. Hao-Song Kong,
Yao Nie, Anthony Vetro, Huifang Sun, Kenneth E. Barner. ICIP 2004.
8
CS MSU Graphics & Media Lab (Video Group)
8. Only for
Maxus
Adaptive Fuzzy Post-Filtering
PSNR 30.61 dB PSNR 30.93 dB
Adaptive Fuzzy Post-Filtering for Highly Compressed Video. Hao-Song Kong,
Yao Nie, Anthony Vetro, Huifang Sun, Kenneth E. Barner. ICIP 2004.
9
CS MSU Graphics & Media Lab (Video Group)
9. Only for
Maxus
Adaptive Fuzzy Post-Filtering
Преимущества
Быстрая работа
Простота в реализации
Недостатки
Просто маскировка артефактов
Замыливание изображения
10
CS MSU Graphics & Media Lab (Video Group)
10. Only for
Maxus
Содержание
Введение
Простые модели
Adaptive Fuzzy Post-Filtering
DCT Re-application
Классификация
Support Vector Regression
Modified Mean-Removed Classified Vector Quantization
Регуляризация
Заключение
11
CS MSU Graphics & Media Lab (Video Group)
11. Only for
Maxus
DCT Re-application
1. Сдвинуть изображение по вертикали и
горизонтали на (i,j)
2. Сжать изображение
3. Вернуть изображение обратно,
т.е. сдвинуть на (-i,-j)
4. Повторить для всех сдвигов из [-3, 4] x [-3, 4]
5. Усреднить результат
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
Richardson. 2002.
12
CS MSU Graphics & Media Lab (Video Group)
12. Only for
Maxus
DCT Re-application
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
13
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
CS MSU Graphics & Media Lab (Video Group) Richardson. 2002.
13. Only for
Maxus
DCT Re-application
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
14
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
CS MSU Graphics & Media Lab (Video Group) Richardson. 2002.
14. Only for
Maxus
DCT Re-application
y i , j D ( i , j )
QD i , j
x
i, j
x - сжатое изображени е
y - восстановл енное изображени е
D - оператор сдвига
Q - оператор кодирования и декодирования
i, j
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
15
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
CS MSU Graphics & Media Lab (Video Group) Richardson. 2002.
15. Only for
Maxus
DCT Re-application
Пример работы
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
Richardson. 2002.
16
CS MSU Graphics & Media Lab (Video Group)
16. Only for
Maxus
DCT Re-application
Пример работы
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
Richardson. 2002.
17
CS MSU Graphics & Media Lab (Video Group)
17. Only for
Maxus
DCT Re-application
Пример работы
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
Richardson. 2002.
18
CS MSU Graphics & Media Lab (Video Group)
18. Only for
Maxus
DCT Re-application
Преимущества
Заметное уменьшение артефактов
Прост в реализации
Недостатки
Желательно знать параметры сжатия
Ориентирован на изображения
Невысокая скорость работы
(~58 умножений на пиксель)
19
CS MSU Graphics & Media Lab (Video Group)
19. Only for
Maxus
Содержание
Введение
Простые модели
Adaptive Fuzzy Post-Filtering
DCT Re-application
Классификация
Support Vector Regression
Modified Mean-Removed Classified Vector Quantization
Регуляризация
Заключение
20
CS MSU Graphics & Media Lab (Video Group)
20. Only for
Maxus
Support Vector Regression
Support Vector Machine
Compression Artifact Reduction Using Support Vector Regression. 21
CS MSU Graphics & Media Lab (Video Group) Sanjeev Kumar, Truong Nguyen, Mainak Biswas. ICIP 2006
21. Only for
Maxus
Support Vector Regression
y f (x) - моделируемая функция
xi , yi - тренировочные данные
y (w), x b - вид модели
Минимизация
w, w C i i*
l
1
2
i 1
при условии
yi w, ( xi ) b i допустимое отклонение для всех элементов
w, ( xi ) b yi i* i , i* отклонение для каждого элемента
, * 0
i i
Compression Artifact Reduction Using Support Vector Regression. 22
CS MSU Graphics & Media Lab (Video Group) Sanjeev Kumar, Truong Nguyen, Mainak Biswas. ICIP 2006
22. Only for
Maxus
Support Vector Regression
Compression Artifact Reduction Using Support Vector Regression. 23
CS MSU Graphics & Media Lab (Video Group) Sanjeev Kumar, Truong Nguyen, Mainak Biswas. ICIP 2006
23. Only for
Maxus
Support Vector Regression
Compression Artifact Reduction Using Support Vector Regression. 24
CS MSU Graphics & Media Lab (Video Group) Sanjeev Kumar, Truong Nguyen, Mainak Biswas. ICIP 2006
24. Only for
Maxus
Support Vector Regression
Compression Artifact Reduction Using Support Vector Regression. 25
CS MSU Graphics & Media Lab (Video Group) Sanjeev Kumar, Truong Nguyen, Mainak Biswas. ICIP 2006
25. Only for
Maxus
Support Vector Regression
Преимущества
Визуальное улучшение на некоторых
типах изображений
Недостатки
Зависимость от обучающей выборки
Сложно прогнозируемое поведение
26
CS MSU Graphics & Media Lab (Video Group)
26. Only for
Maxus
Содержание
Введение
Простые модели
Adaptive Fuzzy Post-Filtering
DCT Re-application
Классификация
Support Vector Regression
Modified Mean-Removed Classified Vector Quantization
Регуляризация
Заключение
27
CS MSU Graphics & Media Lab (Video Group)
27. Only for
Maxus
MMRCVQ (Modified Mean-Removed
Classified Vector Quantization)
Artifact reduction of compressed color images using modified mean-removed classified vector
quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston Lo. Journal of the Chinese Institute
of Engineers, Vol. 27, No. 5, pp. 747-751 (2004)
28
CS MSU Graphics & Media Lab (Video Group)
28. Only for
Maxus MMRCVQ
Построение кодирующих
последовательностей
Выбор блоков 4x4
T ' x'i : i 1,, N сжатые данные
T xi : i 1,, N исходные данные
Нормировка
z ' i x' i x ' i , zi xi xi
Кластеризация
Алгоритм Linde-Buzo-Gray
Artifact reduction of compressed color images using modified mean-removed classified vector
quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston Lo. Journal of the Chinese Institute
of Engineers, Vol. 27, No. 5, pp. 747-751 (2004)
29
CS MSU Graphics & Media Lab (Video Group)
29. Only for
Maxus
MMRCVQ
Linde-Buzo-Gray
Выбор
начального
разбиения
Удвоение числа Оценка
K-means
кластеров дисперсии
Artifact reduction of compressed color images using modified mean-removed classified vector
quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston Lo. Journal of the Chinese Institute
of Engineers, Vol. 27, No. 5, pp. 747-751 (2004)
30
CS MSU Graphics & Media Lab (Video Group)
30. Only for
Maxus
MMRCVQ
Linde-Buzo-Gray
31
CS MSU Graphics & Media Lab (Video Group)
31. Only for
Maxus
MMRCVQ
Linde-Buzo-Gray
32
CS MSU Graphics & Media Lab (Video Group)
32. Only for
Maxus
MMRCVQ
Выбор кодирующих последовательностей
C ' y'i : i 1,..., M
Разбиение несжатых данных
Rk xi | d ( x'i , yk ) d ( x'i , yl ): i l k 1...M , i 1...N
Выбор декодирующих последовательностей
x
x j Ri
j
yi
Ri
Artifact reduction of compressed color images using modified mean-removed classified vector
quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston Lo. Journal of the Chinese Institute
of Engineers, Vol. 27, No. 5, pp. 747-751 (2004)
33
CS MSU Graphics & Media Lab (Video Group)
33. Only for
Maxus
MMRCVQ
Восстановление
-
Классификация
Граница
блока
Определение
Однородный Текстура направления
границы
Отсечение
Поиск плохих Замена блока
результатов
+
Artifact reduction of compressed color images using modified mean-removed classified vector
quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston Lo. Journal of the Chinese Institute
of Engineers, Vol. 27, No. 5, pp. 747-751 (2004)
34
CS MSU Graphics & Media Lab (Video Group)
34. Only for
Maxus
MMRCVQ
Результаты
Artifact reduction of compressed color images using modified mean-removed classified vector
quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston Lo. Journal of the Chinese Institute
of Engineers, Vol. 27, No. 5, pp. 747-751 (2004)
35
CS MSU Graphics & Media Lab (Video Group)
35. Only for
Maxus
MMRCVQ
Результаты
Artifact reduction of compressed color images using modified mean-removed classified vector
quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston Lo. Journal of the Chinese Institute
of Engineers, Vol. 27, No. 5, pp. 747-751 (2004)
36
CS MSU Graphics & Media Lab (Video Group)
36. Only for
Maxus
MMRCVQ
Преимущества
Хорошее качество
Приемлемая скорость восстановления
Недостатки
Только устранение блочности
37
CS MSU Graphics & Media Lab (Video Group)
37. Only for
Maxus
Содержание
Введение
Простые модели
Adaptive Fuzzy Post-Filtering
DCT Re-application
Классификация
Support Vector Regression
Modified Mean-Removed Classified Vector Quantization
Регуляризация
Заключение
38
CS MSU Graphics & Media Lab (Video Group)
38. Only for
Maxus
Regularized Iterative Restoration
f1 g1 n1
f g n f исходный кадр
2 2 2
g наблюдаемы й кадр
f , g , n
fk gk nk n шум
D оператор ллинейно деградации
fN
gN
nN
g Df n
J( f ) f g E ( f )
2
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
39
CS MSU Graphics & Media Lab (Video Group)
39. Only for
Maxus
Regularized Iterative Restoration
Пространственная регуляризация
f (i, j ) f (i 1, j ) if i 8l for l 1,2,..., width/8 - 1
QHB f (i, j )
0 otherwise
f (i, j ) f (i 1, j ) if i 8l for l 1,2,..., width/8 - 1
Q HB
f (i, j )
0 otherwise
QB QHB f QVB f
2 2
2 2
QB QHB f QVB f
Es ( f ) QHB f l
2
QVB f l
2
Q HB
fl
2
QVB f l
2
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
40
CS MSU Graphics & Media Lab (Video Group)
40. Only for
Maxus
Regularized Iterative Restoration
Временная регуляризация
f (i, j ) f i m
width height
2
MFD ( f k , f l ) l k
(i , j )
, j n(i, j ) (i , j )
i 1 j 1
m (i , j )
, n(i, j ) (i , j ) - вектор движения пикселя (i, j )
Et ( f l ) MFD ( f k , f l )
k l
MC 2
Et ( f l ) f l f l
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
41
CS MSU Graphics & Media Lab (Video Group)
41. Only for
Maxus
Regularized Iterative Restoration
Решение
Цель
J ( f ) f g s Es ( f ) t Et ( f )
2
Решение
f i k 1 f i k g I S1 QB QB S2 QB QB t f i f i MC f i k
T T
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
42
CS MSU Graphics & Media Lab (Video Group)
42. Only for
Maxus
Regularized Iterative Restoration
Многокадровое восстановление
~
f f1T , f 2T ,..., f LT T
~ L 2 ~ ~
J ( f ) f l g l s Es ( f ) t Et ( f )
l 1
~ L
l 1
Es ( f ) 1 QHB f l
2
QVB f l
2
Q
2 HB
fl
2
QVB f l
2
L
~
Et ( f ) f l f l MC
2
l 1
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
43
CS MSU Graphics & Media Lab (Video Group)
43. Only for
Maxus
Regularized Iterative Restoration
Многокадровое восстановление
~ ~ ~2 ~2 ~2 ~ ~ MC 2
J ( f ) f g S1 QB f S2 QB f t f f
~ ~ MC
f f
~ k 1 ~ k
f
~
~k
f g I S1 QB QB S 2 QB QB f t
T T
~
f ~ ~k
f f
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
44
CS MSU Graphics & Media Lab (Video Group)
44. Only for
Maxus
Regularized Iterative Restoration
Примеры работы
# of
case
1 spatial only
2 16x16 ME
3 4x4 ME
4 dense 4x4 ME
bidirectional 4x4
5 ME
dense bidirectional
6 4x4 ME
Compressed Video Enhancement using Regularized Iterative Restoration 45
CS MSU Graphics & Media Lab (Video Group) Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
45. Only for
Maxus
Regularized Iterative Restoration
Примеры работы
dense 4x4 ME
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
46
CS MSU Graphics & Media Lab (Video Group)
46. Only for
Maxus
Regularized Iterative Restoration
Примеры работы
dense bidirectional 4x4 ME
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
47
CS MSU Graphics & Media Lab (Video Group)
47. Only for
Maxus
Regularized Iterative Restoration
Оптимизация
2
J ( f ) f g s QB f t f f MC
2
- цель минимизации
J ( f ) f g s QBi , j f t f f MC
2
i, j
идея: добавить вспомогательную переменную, не изменяющую минимум J
t inf bt 2 b
b
' t
bt - дополнительная переменная
2t
J * ( f , b) f g s bi , j QBi , j f
2
b
2
t f f MC
i, j
- новая цель минимизации
A Coding Artifacts Removal Algorithm Based On Spatial And Temporal
Regularization. Susu Yao, Genan Feng, Xiao Lin, Keng Pang Lim and
Weisi Lin. 2003. 48
CS MSU Graphics & Media Lab (Video Group)
48. Only for
Maxus
Regularized Iterative Restoration
Оптимизация
Алгоритм:
Initializa tion : f 0 g
repeat {
b n 1 min J * f n , b n with f n fixed :
b n 1 QB f n f n
f n 1 min J * f n , b n 1 with b fixed
} Until convergence
Минимизация может быть выполнена методом градиентного спуска
A Coding Artifacts Removal Algorithm Based On Spatial And Temporal
Regularization. Susu Yao, Genan Feng, Xiao Lin, Keng Pang Lim and
Weisi Lin. 2003. 49
CS MSU Graphics & Media Lab (Video Group)
49. Only for
Maxus
Regularized Iterative Restoration
Оптимизация
i, j
2
i, j
j 2
j
i, j
J * ( f n , b n1 ) f i ,nj f i ,nj1 s bin, 1 QB i , j f n1 bin, 1 t f i ,nj1 f i ,MC
j 2
s bin, 1
n 1
fi, j
1
1 t
xi , j
n j
1 t 1 2sbin, 1 t
fi,nj fi,nj 1
j
t 1 s bin, 1 t s t bin, 1
f i ,mc
j j
f i ,mc 1
1 t 1 2sbin, 1 t j 1 t 1 2sbin, 1 t j
j j
A Coding Artifacts Removal Algorithm Based On Spatial And Temporal
Regularization. Susu Yao, Genan Feng, Xiao Lin, Keng Pang Lim and
Weisi Lin. 2003. 50
CS MSU Graphics & Media Lab (Video Group)
50. Only for
Maxus
Regularized Iterative Restoration
Пример работы
A Coding Artifacts Removal Algorithm Based On Spatial And Temporal
Regularization. Susu Yao, Genan Feng, Xiao Lin, Keng Pang Lim and 51
CS MSU Graphics & Media Lab (Video Group) Weisi Lin. 2003.
51. Only for
Maxus
Regularized Iterative Restoration
Пример работы
A Coding Artifacts Removal Algorithm Based On Spatial And Temporal
Regularization. Susu Yao, Genan Feng, Xiao Lin, Keng Pang Lim and Weisi 52
CS MSU Graphics & Media Lab (Video Group) Lin. 2003.
52. Only for
Maxus
Regularized Iterative Restoration
Пример работы
A Coding Artifacts Removal Algorithm Based On Spatial And Temporal
Regularization. Susu Yao, Genan Feng, Xiao Lin, Keng Pang Lim and
Weisi Lin. 2003. 53
CS MSU Graphics & Media Lab (Video Group)
53. Only for
Maxus
Regularized Iterative Restoration
Преимущества
Высокое качество восстановления
Недостатки
Очень медленная работа
Сложен в реализации
54
CS MSU Graphics & Media Lab (Video Group)
54. Only for
Maxus
Заключение
Рассмотрены методы
Adaptive Fuzzy Post-Filtering
DCT Re-application
Support Vector Regression
Modified Mean-Removed Classified Vector Quantization
Регуляризация
55
CS MSU Graphics & Media Lab (Video Group)
55. Only for
Maxus
Литература
• A Coding Artifacts Removal Algorithm Based On Spatial And Temporal
Regularization. Susu Yao, Genan Feng, Xiao Lin, Keng Pang Lim and Weisi Lin.
2003.
Artifact reduction of compressed color images using modified mean-removed
classified vector quantization. Jim Zone-Chang Lai, Yi-Ching Liaw, and Winston
Lo. Journal of the Chinese Institute of Engineers, Vol. 27, No. 5, pp. 747-751
(2004)
Compressed Video Enhancement using Regularized Iterative Restoration
Algorithms. Passant Vatsalya Karunaratne. Evanston, Illinois. 1999
Compression Artifact Reduction Using Support Vector Regression. Sanjeev
Kumar, Truong Nguyen, Mainak Biswas. ICIP 2006
Adaptive Fuzzy Post-Filtering for Highly Compressed Video. Hao-Song Kong, Yao
Nie, Anthony Vetro, Huifang Sun, Kenneth E. Barner. ICIP 2004.
Enhancement of JPEG-Compressed Images by Re-application of JPEG. Aria
Nosratinia. Department of Electrical Engineering, University of Texas at Dallas,
Richardson. 2002.
A Tutorial on Support Vector Regression. Alex J. Smola† and Bernhard
Sch¨olkop. 2003
Application Of The Motion Vector Constraint To The Regularized Enhancement Of
Compressed Video. C. Andrew Segall And Aggelos K. Katsaggelos. 2001
56
CS MSU Graphics & Media Lab (Video Group)
56. Only for
Maxus
Вопросы
?
57
CS MSU Graphics & Media Lab (Video Group)