Lidu - Revision history

imported>Projects221: /* Final Workflow */

2013-03-21T04:30:06Z

Final Workflow

← Older revision		Revision as of 04:30, 21 March 2013
Line 97:		Line 97:

	=== Final Workflow ===		=== Final Workflow ===
	Since different methods can be used to detect doubly compressed images with Q1<Q2 and Q1>Q2, so my detection can be divided into three parts. The first part is to use simple histogram methods to detect doubly compressed image with Q1<Q2. Then if the result is not true, the image could be singly compressed or doubly compressed with Q1>Q2. The second part is to use SVM or Naive Bayes to do further classification. If the result is true, the image is doubly compressed, ~~other~~ in the third part, it is classified as the singly compressed image.		Since different methods can be used to detect doubly compressed images with Q1<Q2 and Q1>Q2, so my detection can be divided into three parts.
			# The first part is to use simple histogram methods to detect doubly compressed image with Q1<Q2.
			# Then if the result is not true, the image could be singly compressed or doubly compressed with Q1>Q2. The second part is to use SVM or Naive Bayes to do further classification.
			# If the result is true, the image is doubly compressed, otherwise in the third part, it is classified as the singly compressed image.

	== Results ==		== Results ==

imported>Projects221: /* Naive Bayes */

2013-03-21T04:27:53Z

Naive Bayes

← Older revision		Revision as of 04:27, 21 March 2013
Line 94:		Line 94:
	<br>		<br>
	[[File:prob.jpg \| center\|400px \| math 6]]		[[File:prob.jpg \| center\|400px \| math 6]]
	[[File:pred.jpg \| center\|~~400px~~ \| math 7]]		[[File:pred.jpg \| center\|350px \| math 7]]

	=== Final Workflow ===		=== Final Workflow ===

imported>Projects221: /* Naive Bayes */

2013-03-21T04:27:11Z

Naive Bayes

← Older revision		Revision as of 04:27, 21 March 2013
Line 94:		Line 94:
	<br>		<br>
	[[File:prob.jpg \| center\|400px \| math 6]]		[[File:prob.jpg \| center\|400px \| math 6]]
			[[File:pred.jpg \| center\|400px \| math 7]]

	=== Final Workflow ===		=== Final Workflow ===

imported>Projects221: /* Markov Random Process Prediction */

2013-03-21T04:25:24Z

Markov Random Process Prediction

← Older revision		Revision as of 04:25, 21 March 2013
Line 82:		Line 82:
	For example, m = n = -4, the denominator is the total number of -4 in this array, and the numerator is the number of pairs in which the current value is -4, the next value (in the horizontal direction) is also -4. So the result of p{next = -4 \| current = -4} is the transition probability of -4 to -4. Do the similar calculation for all current value and next value in the range of [-4, 4], so there would be 99 = 81 probabilities in one direction, and there would be 99*4 probabilities in all directions. These probabilities can be used as features by machine learning methods.		For example, m = n = -4, the denominator is the total number of -4 in this array, and the numerator is the number of pairs in which the current value is -4, the next value (in the horizontal direction) is also -4. So the result of p{next = -4 \| current = -4} is the transition probability of -4 to -4. Do the similar calculation for all current value and next value in the range of [-4, 4], so there would be 99 = 81 probabilities in one direction, and there would be 99*4 probabilities in all directions. These probabilities can be used as features by machine learning methods.

	However, computation of this method is too large. For one image, 81 iterations are need to calculate the transition probabilities, and in order to use machine learning methods, 1000 or more images are needed to do the training, so this method is very slow and inefficient.		However, computation of this method is too large. For one image, 81 iterations are need to calculate the transition probabilities, and in order to use machine learning methods, 1000 or more images are needed to do the training, so this method is very slow and inefficient, and would be not developed further in this project.

	=== Machine Learning ===		=== Machine Learning ===

imported>Projects221: /* Markov Random Process Prediction */

2013-03-21T04:24:42Z

Markov Random Process Prediction

← Older revision		Revision as of 04:24, 21 March 2013
Line 79:		Line 79:
	[[File:transition.jpg \| center \| math 5]]		[[File:transition.jpg \| center \| math 5]]

			δ(x = a) is one if x == a, otherwise, it is 0.
	For example, m = n = -4, the denominator is the total number of -4 in this array, and the numerator is the number of pairs in which the current value is -4, the next value (in the horizontal direction) is also -4. So the result of p{next = -4 \| current = -4} is the transition probability of -4 to -4. Do the similar calculation for all current value and next value in the range of [-4, 4], so there would be 99 = 81 probabilities in one direction, and there would be 99*4 probabilities in all directions. These probabilities can be used as features by machine learning methods.		For example, m = n = -4, the denominator is the total number of -4 in this array, and the numerator is the number of pairs in which the current value is -4, the next value (in the horizontal direction) is also -4. So the result of p{next = -4 \| current = -4} is the transition probability of -4 to -4. Do the similar calculation for all current value and next value in the range of [-4, 4], so there would be 99 = 81 probabilities in one direction, and there would be 99*4 probabilities in all directions. These probabilities can be used as features by machine learning methods.

imported>Projects221: /* AC Coefficients Histogram */

2013-03-21T04:23:17Z

AC Coefficients Histogram

← Older revision		Revision as of 04:23, 21 March 2013
Line 24:		Line 24:
	=== Histograms ===		=== Histograms ===
	==== AC Coefficients Histogram ====		==== AC Coefficients Histogram ====
	Usually, in the process of double compression, the two quantization steps are different, which may cause a different histogram distribution in doubly compressed images. Some people did research on the distribution of AC coefficients, and found that the distributions satisfy a Laplacian distribution, which is centered at 0, and decreases in each side. Furthermore, the distribution of the difference of nearby values also satisfy a Laplacian distribution. So this method applies this pattern to detect double compression.Since the upperleft coefficients contains the information of low frequencies, they are more sensitive to high frequencies, I choose the first 20 modes in zig-zag order, and draw the histogram of values from 2 to 10 in these modes. That is, in each mode for all blocks, I record the number of appearance of 2, 3, 4, ..., 10, so there are 9 buckets in each mode. The difference histogram calculates the difference of counts of nearby values, that mean, the first bucket stores f(2)-f(3), the second bucket stores f(3)-f(4)..., where f(x) means the number of appearance. Here are two difference histogram examples, the left one is the histogram of a singly compressed image with quality factor 80, and right one is of a doubly compressed image with quality 65 and 95 separately.		Usually, in the process of double compression, the two quantization steps are different, which may cause a different histogram distribution in doubly compressed images. Some people did research on the distribution of AC coefficients, and found that the distributions satisfy a Laplacian distribution, which is centered at 0, and decreases in each side. Furthermore, the distribution of the difference of nearby values also satisfy a Laplacian distribution. So this method applies this pattern to detect double compression.Since the upperleft coefficients contains the information of low frequencies, they are more sensitive to high frequencies, I choose the first 20 modes in zig-zag order, and draw the histogram of values from 2 to 10 in these modes. That is, in each mode for all blocks, I record the number of appearance of 2, 3, 4, ..., 10, so there are 9 buckets in each mode. The difference histogram calculates the difference of counts of nearby values, that mean, the first bucket stores f(2)-f(3), the second bucket stores f(3)-f(4)..., where f(x) means the number of appearance of value x in one mode. Here are two difference histogram examples, the left one is the histogram of a singly compressed image with quality factor 80, and right one is of a doubly compressed image with quality 65 and 95 separately.
	<br>		<br>
	[[File: single.jpg \|400px \| Figure 1]] [[File: double.jpg \|400px \| Figure 2]]		[[File: single.jpg \|400px \| Figure 1]] [[File: double.jpg \|400px \| Figure 2]]

imported>Projects221: /* Double JPEG Compression */

2013-03-21T04:22:23Z

Double JPEG Compression

← Older revision		Revision as of 04:22, 21 March 2013
Line 18:		Line 18:
	<br>		<br>
	[[File:doubleflow.jpg\| center \| 500px \| Figure 2]]		[[File:doubleflow.jpg\| center \| 500px \| Figure 2]]

			Corresponding to the quantization step, another term is quality factor. The bigger the quality factor is, the better the image is, the smaller the quantization step is. In this project, for double compression, the quality factor for the first compression is named Q1 and that for the second compression is named Q2.

	== Method ==		== Method ==

imported>Projects221: /* Background */

2013-03-21T04:19:36Z

Background

← Older revision		Revision as of 04:19, 21 March 2013
Line 1:		Line 1:
	== Background ==		== Background ==
	With the development of digital image processing in recent ~~years，~~ numerous image editing techniques have been developed so that realistic synthetic images can be produced conveniently without leaving noticeable visual artifacts. Although these image editing technologies can enrich the user experience, they also cause the problem for people to trust the authenticity of the images. JPEG is the most popular image format for cameras, and the modification of JPEG images may be re-saved to generate a new image, which would cause the double compression of the image. This project focuses on the detection of double compression and single compression to verify the authenticity of the image.		With the development of digital image processing in recent years, numerous image editing techniques have been developed so that realistic synthetic images can be produced conveniently without leaving noticeable visual artifacts. Although these image editing technologies can enrich the user experience, they also cause the problem for people to trust the authenticity of the images. JPEG is the most popular image format for cameras, and the modification of JPEG images may be re-saved to generate a new image, which would cause the double compression of the image. This project focuses on the detection of double compression and single compression to verify the authenticity of the image.

	== Introduction ==		== Introduction ==

imported>Projects221: /* AC Coefficients Histogram */

2013-03-20T22:43:17Z

AC Coefficients Histogram

← Older revision		Revision as of 22:43, 20 March 2013
Line 22:		Line 22:
	=== Histograms ===		=== Histograms ===
	==== AC Coefficients Histogram ====		==== AC Coefficients Histogram ====
	Usually, in the process of double compression, the two quantization steps are different, which may cause a different histogram distribution in doubly compressed images. Some people did research on the distribution of AC coefficients, and found that the distributions satisfy a Laplacian distribution, which is centered at 0, and decreases in each side. Furthermore, the distribution of the difference of nearby values also satisfy a Laplacian distribution. So this method applies this pattern to detect double compression.Since the upperleft coefficients contains the information of low frequencies, they are more sensitive to high frequencies, I choose the first 20 modes in zig-zag order, and draw the histogram of values from 2 to 10 in these modes. Here are two examples, the left one is the histogram of a singly compressed image with quality factor 80, and right one is of a doubly compressed image with quality 65 and 95 separately.		Usually, in the process of double compression, the two quantization steps are different, which may cause a different histogram distribution in doubly compressed images. Some people did research on the distribution of AC coefficients, and found that the distributions satisfy a Laplacian distribution, which is centered at 0, and decreases in each side. Furthermore, the distribution of the difference of nearby values also satisfy a Laplacian distribution. So this method applies this pattern to detect double compression.Since the upperleft coefficients contains the information of low frequencies, they are more sensitive to high frequencies, I choose the first 20 modes in zig-zag order, and draw the histogram of values from 2 to 10 in these modes. That is, in each mode for all blocks, I record the number of appearance of 2, 3, 4, ..., 10, so there are 9 buckets in each mode. The difference histogram calculates the difference of counts of nearby values, that mean, the first bucket stores f(2)-f(3), the second bucket stores f(3)-f(4)..., where f(x) means the number of appearance. Here are two difference histogram examples, the left one is the histogram of a singly compressed image with quality factor 80, and right one is of a doubly compressed image with quality 65 and 95 separately.
	<br>		<br>
	[[File: single.jpg \|400px \| Figure 1]] [[File: double.jpg \|400px \| Figure 2]]		[[File: single.jpg \|400px \| Figure 1]] [[File: double.jpg \|400px \| Figure 2]]

imported>Projects221: /* Machine Learning */

2013-03-20T22:37:25Z

Machine Learning

← Older revision		Revision as of 22:37, 20 March 2013
Line 91:		Line 91:
	<br>		<br>
	[[File:prob.jpg \| center\|400px \| math 6]]		[[File:prob.jpg \| center\|400px \| math 6]]

			=== Final Workflow ===
			Since different methods can be used to detect doubly compressed images with Q1<Q2 and Q1>Q2, so my detection can be divided into three parts. The first part is to use simple histogram methods to detect doubly compressed image with Q1<Q2. Then if the result is not true, the image could be singly compressed or doubly compressed with Q1>Q2. The second part is to use SVM or Naive Bayes to do further classification. If the result is true, the image is doubly compressed, other in the third part, it is classified as the singly compressed image.

	== Results ==		== Results ==