Image authentic identification method
By integrating traditional ELA error level analysis method and deep learning CNN model, preprocessing and feature adaptive learning are performed on images, solving the problem of low accuracy in the prior art, and effectively detecting fine-grained features and various tampering methods.
Patent Information
- Application Number
- CN202510118433.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-24
- Publication Date
- 2025-06-20
AI Technical Summary
The existing image pseudo-recognition technology has low accuracy, and it is especially difficult to effectively detect micro-editing tampering and fine-grained features.
The traditional methods and deep learning methods are integrated, and the images are preprocessed using traditional image pseudo-detection ELA error level analysis method, and feature adaptive learning is performed in combination with deep learning methods such as CNN models to realize binary classification judgment of images.
It significantly improves the ability to identify fine-grained features, improves the accuracy and robustness of image tamper detection, and can better detect multiple tampering methods.
Smart Images
Figure CN120182797A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of digital image forensics, and particularly to an image forgery detection method that combines traditional methods and deep learning. Background Art
[0002] Image forgery detection technology provides technical support for fields such as court litigation, news media, medicine, military, and scientific research, and has important research significance for social development and people's lives. However, with the development of information technology, multimedia information such as digital images is almost everywhere in the Internet and the real world. Facing various forgery means brought about by technological progress, the original image is easily maliciously tampered with without leaving any obvious traces. With the frequent occurrence of image forgery incidents, image forgery detection is urgent.
[0003] The main ways of image tampering are splicing, copy-pasting, and deletion. These three operations modify the image content, and the modification methods are highly misleading and are also the three most common tampering methods. The remaining modification methods such as blurring, compression, enhancement, scaling, filtering, etc. do not modify the image content. Most of them are post-processing operations to cover up the tampering traces of the image. Commonly, image tampering forensics technology is divided into two categories: traditional methods and deep learning methods.
[0004] Traditional methods: First, feature extraction is performed by manually designing features, and then the features are compared and analyzed or the statistical characteristics of the image are analyzed to determine whether the image has been tampered with. Most traditional methods can only target one type of image tampering method. By manually extracting features, building models, and classifying features, the classification results are finally obtained to determine whether the image has been tampered with.
[0005] Deep learning: The adaptive feature extraction method is a feature extraction method that best suits the deep learning idea. Usually, after the image is input, no artificial feature extraction operation is performed, and the neural network automatically learns the tampering features in the image.
[0006] Currently, there is no method in traditional methods that can deal with all image tampering techniques. Most of them can only target one type of image tampering method. By manually designing feature extraction, building a model, and classifying features, the classification result is finally obtained to judge whether the image has been tampered with. With the successful application of deep learning technology in various computer vision and image processing tasks, in recent years, some researchers have carried out image forgery forensics through deep learning methods and achieved good results in the identification of various types of tampering techniques, bringing hope for a general image tampering forensics method. The algorithms of deep learning mainly use the CNN model to extract features and classify through softmax to achieve an end-to-end adaptive learning mode. The biggest advantage of this end-to-end automatic detection system is to automatically learn feature parameters and adapt to the detection of various tampering methods. In addition, the features obtained by deep learning have stronger expressive power and can learn the fine-grained features of the image, that is, they can better learn the tampering features of the image.
[0007] Currently, most image forgery detections use traditional methods or deep learning methods. Traditional methods require extracting manually designed features and then judging whether the image has been tampered with based on experience or training a classifier. Supervised learning in deep learning is to learn images and labels through a network, then extract features of different categories, and finally judge whether the image has been tampered with. The inventor found during the research that it is easy to distinguish whether some images have been tampered with by the features extracted by traditional methods. However, this method is limited by subjective human consciousness. Especially for images with minor editing tampering, it is very difficult to define whether they have been tampered with, and some fine-grained features cannot be effectively extracted and detected. Summary of the Invention
[0008] To solve the problem of low accuracy in existing image forgery detection technologies, the present invention proposes an image forgery detection method. This method combines traditional methods and deep learning methods, significantly improving the ability to identify fine-grained features and better solving the problem of image tampering.
[0009] The specific technical solution is as follows:
[0010] An image forgery detection method includes the integration of traditional image forgery detection methods and deep learning methods. The traditional image forgery ELA error level analysis method is used to preprocess the image, and the processed result is input into the deep learning method for adaptive learning of features to achieve binary classification of tampered and untampered.
[0011] Furthermore, the traditional image forgery detection method also includes the brightness gradient method and the spatial color method.
[0012] Furthermore, the 2D histogram is drawn using the image chromaticity and saturation of HSV and two chromaticities of Lab to judge whether the image has been tampered with.
[0013] Furthermore, the image is processed by using ELA, brightness gradient method, and spatial color method respectively, and the processed images are fused to achieve preprocessing.
[0014] Furthermore, the deep learning methods include CNN, VGG, ResNet, Inception, and Xception.
[0015] Beneficial effects
[0016] 1. The boundary for judging whether an image is tampered with by traditional methods is limited by human subjective consciousness. Especially for images with minor editing tampering, it is very difficult to determine whether they are tampered with. Using deep learning methods, namely convolutional neural networks, to automatically learn tampering features can ultimately determine whether the image is tampered with, but the robustness is weak. The inventor further studies and finds that the method of combining traditional methods and deep learning can more accurately judge whether an image is tampered with, and the effect is the best.
[0017] 2. In this era of rapid development of science and technology, various digital devices are developing rapidly, such as mobile phones, computers, cameras, etc. These tools are constantly pouring into our lives. We can pick up our mobile phones to freeze the beautiful moments in life. Professional technicians can use various advanced cameras to complete the work in this field. People can also perform fine image processing through application software on the computer. It is not difficult to see that various changes in digital images have spread to every corner of our daily lives.
[0018] The digital image forgery detection technology deeply studied in this application is inspired by this aspect. With the continuous innovation of mobile phone and tablet software, image tampering can not only be completed on professional systems and devices, but every ordinary person can beautify their photos through various "photo beautification software". On the one hand, positive image tampering can make our photos more perfect and pleasing to the eye; but on the other hand, there are many lawbreakers who use various means to forge pictures to pass off the false as the real. Once they flow into society, with the rapid spread of the Internet, a small picture may have extremely bad effects. It may not only have a huge impact on an individual's reputation and interests, but also indirectly have an adverse impact on social stability and unity, and even may have a huge impact on national security.
[0019] It can be seen that while doing a good job in digital image forgery detection, the present invention is a measure that benefits and eliminates disadvantages for maintaining social order and safeguarding national and corporate security.
[0020] 3. The present invention uses a variety of traditional methods to preprocess the image, adds prior information (i.e., highlighting factors) to the data using manual features to make the data more separable, and then inputs it into a deep neural network for model training, so as to better detect whether the image is tampered with. Brief Description of the Drawings
[0021] Figure 1 、Flowchart of the method of the present invention. Detailed Embodiments
[0022] The key contents involved in the present invention will be elaborated in detail below.
[0023] Key Point 1: Traditional Method
[0024] (1) ELA Error Level Analysis Method
[0025] ELA is a relatively classic image forgery forensics algorithm. The full English name of ELA is "Error Level Analysis", which is translated as "Error Level Analysis" or "Error Analysis". It judges whether an image has been processed such as PS splicing, modification, or smearing by detecting the error distribution caused by redrawing the image after the compression ratio and analyzing the compression ratio of each region of the JPEG image and calculating the error level. Generally speaking, it allows to identify regions with different compression levels of image processing, highlighting the differences in JPEG compression ratios. Regions with uniform coloring may have lower ELA values than high-contrast edges. In principle, for the original JPEG image obtained by only one sampling, the ELA values of each region should be similar. If the ELA value of a region is significantly different from other parts of the image, then this region is very likely to be modified later.
[0026] Specifically, ELA divides the picture into many squares and performs a separate color space conversion on each small block. Every time the JPEG image is modified, a second conversion will be performed. Naturally, there will be differences between the two conversions, and ELA judges which part of the picture has been modified by comparing this difference. ELA detection can be divided into the following three parts:
[0027] ①Points refer to repeated textures or similar data in the picture. Repeated textures should show approximate colors during ELA analysis, and the data differences in regions with more details should also be large.
[0028] ②Lines are the boundary lines between large areas of different colors. The same contrast edges should show approximate ELA results. The greater the contrast, the higher the ELA value and the clearer the line.
[0029] ③For surfaces, there are no differences in solid color surfaces, so there is no ELA, only black or black coloring.
[0030] (2) Brightness Gradient Method.
[0031] Specifically, the principle is to calculate the gradient value of brightness. Image brightness refers to the light and dark degree of the image. If the pixel value is within [0, 255], the closer the pixel value is to 0, the lower the brightness. An edge is a place where the pixel value changes rapidly.
[0032] The image brightness gradient represents the speed of image change and reflects the edge information of the image. For the edge part of the image, its brightness value changes greatly, and the gradient value is also large; for the smoother part of the image, its brightness value changes little, and the gradient value is also small.
[0033] In principle, the original image should contain discontinuous noise and jagged line segments. If there are smooth blurs or straight line segments in the figure, it is likely the result of image editing.
[0034] (3) Spatial color method.
[0035] Specifically, the spatial color method is to obtain the brightness information of the image. This application uses a color histogram to analyze the spatial color of the image. A color histogram is the proportion of different colors in the entire image, regardless of the spatial position of each color.
[0036] More specifically, RGB color space: The model is easy to understand, but it is not intuitive when continuously changing colors. The RGB (Red, Green, Blue) primary colors all have a value range of [0, 255], [0, 255], [0, 255].
[0037] HSV color space: In order to digitize the image, it cannot well represent the process of the human eye interpreting the image. H (Hue): [0, 360]; S (Saturation), that is, color purity, 0 saturation is white; V (Value / Brightness): Brightness, 0 brightness is pure black. In OpenCV, the color range is: H = [0, 179], S = [0, 255], V = [0, 255]
[0038] Lab color space: The Euclidean distance between colors has a specific meaning - the greater the distance, the greater the perceived difference between the two colors by the human eye. L channel: Pixel brightness, white at the top, black at the bottom, and gray in the middle; a channel: green on the left, red on the right; b channel: pure blue at one end and pure yellow at the other end
[0039] Gray color space: Grayscale image, each pixel is [0, 255]. To convert an RGB image to a grayscale image according to the human eye sensitivity, it is not simply to take the average value of each RGB channel, but: Y = 0.299 * R + 0.587 * G + 0.114 * B.
[0040] Generally speaking, the histogram of an image will be spread out in most color spaces, and the histogram distribution of an edited (color-transformed) image is generally compressed.
[0041] Calculating the image color histogram has a relatively low cost and has many advantages such as invariance to image translation, rotation, and scaling. The color histogram of a general image is a one-dimensional statistical analysis of image brightness, which only considers one characteristic, that is, the gray value of pixels. The present invention delves into the image color space and simultaneously considers two features, namely the chromaticity and saturation of the HSV image and the two chromaticities of Lab, to draw a 2D histogram for analyzing and determining whether the image has been tampered with.
[0042] Key Point Two: Deep learning technology
[0043] With the successful application of deep learning technology in various computer vision and image processing tasks, in recent years, some researchers have used deep learning methods for image forgery forensics and achieved good results in the identification of various forgery techniques. Deep learning methods can regard image forgery forensics as a target detection problem or an anomaly detection problem, bringing hope for general image forgery forensics methods. The algorithms of deep learning mainly use the CNN model to extract features and implement an end-to-end adaptive learning mode. The greatest advantage of this end-to-end automatic detection system is that it can automatically learn feature parameters and is suitable for the detection of various forgery methods.
[0044] The CVPR2019 paper (ManTra-Net: Manipulation Tracing Network For Detection And Localiztion) based on abnormal features to locate the forgery traces in images found that VGG is the best through the verification of the forgery classification network. The VGG network is a lesson from the past. Of course, it has been verified that this network does have good results in extracting forged images. Therefore, other networks such as ResNet, Inception, and Xception that have emerged later can also be selected.
[0045] Key Technology Three: The integration of traditional methods and deep learning
[0046] Traditional image forgery forensics methods have defects. The forensics methods use a manually designed method to extract features. Most of these manually designed features have limitations and lack representativeness, and it is impossible to determine various forgery methods simultaneously based on these features. This has led to the fact that image forensics methods only have the ability to identify one forgery technique.
[0047] In addition, the features based on manual design can currently achieve good results in terms of accuracy, but due to the rapid development of tampering technology, as the means of tampering become more and more advanced and complex, the features based on manual design are easily attacked and blocked by one or several tampering technologies. In other words, we cannot simply rely on ELA to make judgments. Sometimes this method is useless. This method cannot be used in other unknown fields such as photo reshoots, CG painting, and photo color processing.
[0048] Furthermore, the current methods for adaptively extracting image tampering features are not mature enough. It is difficult to achieve robust results by relying solely on regional edge anomalies and intrinsic statistical features of the image. Edge anomalies and statistical features are easily destroyed by the image post-processing process. The constrained convolution layer is a good direction for general feature extraction, but the currently implemented constrained convolution layer can be regarded as a general high-pass filter, which works well in operations such as image enhancement and scaling, but is equivalent to the manually designed high-pass filter in passive image content forensics, and has not achieved better results.
[0049] Therefore, inspired by traditional methods and deep learning to detect tampered images, this application combines ELA with deep learning methods to build a CNN network for classification of the original image and the tampered image after ELA (error-level analysis). However, it is not limited to using ELA, and the image can be processed using traditional methods and then input into the neural network for more in-depth learning.
[0050] The present invention relates to the technical field of digital image forensics, and more specifically, to an image authentication method that integrates traditional methods and deep learning.
[0051] This design is an image authentication method, and this embodiment includes a traditional method, a deep learning method, and a fusion of the two.
[0052] 1. The following are several traditional methods for image authentication:
[0053] This embodiment adopts
[0054] (1) ELA error level analysis method: If the ELA value of a region is significantly different from that of other parts of the image, it means that this region is likely to have been modified later, possibly through splicing, modification, smearing, etc.
[0055] (2) Brightness gradient method: Use edge detection algorithms such as Canny, Laplacian, Sobel, and Schaar to analyze whether the image has been tampered with. In principle, the original image should contain discontinuous noise and jagged line segments. If smooth blur or straight line segments appear in the image, it may be a tampered image.
[0056] (3) Spatial color method: The RGB color space, HSV color space, and Lab color space are used for image forgery detection. In addition, in this embodiment, the image color space is further explored, and two features are considered simultaneously, namely the image chromaticity and saturation of HSV and two chromaticities of Lab, to draw a 2D histogram for analyzing and determining whether the image has been tampered with. Generally, the color distribution of the image is displayed in the form of a histogram, and the histogram of the image will be tiled in most color spaces. The histogram distribution of the tampered image is generally compressed.
[0057] 2. Image Forgery Detection Based on Deep Learning Method
[0058] There are still defects in traditional image forgery forensics methods. These image forensics methods usually extract features in a manually designed manner. Most of these manually designed features have limitations and lack representativeness, and it is impossible to determine multiple tampering methods simultaneously based on these features. This has led to the fact that image forensics methods only have the ability to identify one tampering technique.
[0059] Currently, the features based on manual design can achieve good results in terms of accuracy. However, due to the rapid development of tampering techniques, as the tampering means become more and more advanced and complex, the features based on manual design are easily attacked and blocked by one or several tampering techniques. That is, one cannot simply rely on ELA to judge. Sometimes this method is useless and cannot be used in other unknown fields such as the "Zhou Tiger" (photo reshooting), CG painting, and photo color adjustment processing.
[0060] With the successful application of deep learning technology in various computer vision and image processing tasks, in recent years, some researchers have used deep learning methods for image forgery forensics and achieved good results in the identification of various types of tampering techniques. The deep learning method can regard image forgery forensics as an object detection problem or an anomaly detection problem, bringing hope to the general image forgery forensics method. The algorithm of deep learning mainly uses the CNN model to extract features and classifies through softmax to achieve an end-to-end adaptive learning mode. The greatest advantage of this end-to-end automatic detection system is to automatically learn feature parameters and adapt to the detection of multiple tampering methods.
[0061] In the embodiment, two deep neural networks, VGG16 and VGG19, are used to build a feature extraction model for authenticating forged images. The CVPR2019 paper (ManTra-Net: Manipulation Tracing Network For Detection And Localiztion) verifies that VGG is the best for the tampering classification network based on abnormal features. Therefore, VGG is selected, and a series of experiments are also conducted on the use of different convolutional kernels. Finally, the best combined convolutional features are selected. The VGG network is a lesson from the past. Of course, it is verified that this network does extract better results on tampered images. Therefore, other networks such as ResNet, Inception, and Xception that may rise to the top later can also be selected.
[0062] 3. Image Forgery Detection by Combining Traditional Methods and Deep Learning Methods
[0063] The current method for adaptively extracting image tampering features is still not well-developed. It is very difficult to obtain good robustness results relying solely on regional edge anomalies and intrinsic statistical features of images. Both edge anomalies and statistical features are easily destroyed during the image post-processing process. The constrained convolutional layer is a good direction for general feature extraction. However, the currently implemented constrained convolutional layer can be regarded as a general high-pass filter, which has good effects in operations such as image enhancement and scaling. But in passive image forensics, its effect is equivalent to that of a manually designed high-pass filter, and no better results have been achieved.
[0064] Inspired by traditional methods for detecting tampered images, in this embodiment, ELA is combined with deep learning methods to build a CNN network for classifying the results of original genuine and forged images after ELA (error-level analysis). This neural network will determine whether the image has been modified and give the final probability result of being modified.
[0065] The present invention can preprocess the image using a single traditional image forgery detection method alone, or can use multiple traditional image forgery detection methods to preprocess the image simultaneously, and then send the information after preprocessing to the deep learning method for recognition.
[0066] In summary, the above are only the preferred embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. An image authentication method, characterized in that: The traditional image authentication method is integrated with the deep learning method. The traditional image authentication ELA error level analysis method is used to pre-process the image, and the processed results are input into the deep learning method for adaptive feature learning to achieve binary classification of tampered and non-tampered images.
2. The image authentication method according to claim 1, characterized in that: Traditional image authentication methods also include brightness gradient method and spatial color method.
3. The image authentication method according to claim 2, characterized in that: The image hue and saturation of HSV and the two hues of Lab are used to draw a 2D histogram to determine whether the image has been tampered with.
4. An image authentication method according to any one of claims 1 to 3, characterized in that: The images are processed using ELA, brightness gradient method and spatial color method respectively, and the processed images are fused to achieve preprocessing.
5. The image authentication method according to claim 1, characterized in that: Deep learning methods include CNN, VGG, ResNet, Inception, and Xception.