CN113012152B - Image tampering chain detection method, device and electronic equipment - Google Patents

Abstract

The invention discloses an image tampering chain detection method, an image tampering chain detection device and electronic equipment, wherein the method comprises the following steps: inputting an image to be detected into a preset image feature extractor for feature extraction to obtain a target feature map corresponding to the image to be detected; performing convolution operation on the target characteristic graph and convolution kernels of the target number, and flattening a convolution result to obtain word vector representation of target dimensionality; and inputting the word vector representation into a preset image processing algorithm translation model for image processing algorithm recognition to obtain a tampering chain detection result of the image to be detected, wherein the preset image processing algorithm translation model is constructed on the basis of an attention mechanism model.

Description

Image tampering chain detection method, device and electronic equipment

technical field

The invention relates to the technical field of image detection, in particular to an image tampering chain detection method, device and electronic equipment.

Background technique

Manipulation-Chain refers to the tampering history of an image, that is, the orderly combination of a limited number of image processing algorithms that the image has undergone since its generation. Image processing algorithms generally include: JPEG compression, median filtering, Gaussian blur, super-resolution, etc. , if an image is first compressed by JPEG, then processed by median filter, and finally processed by Gaussian blur, the corresponding tampering chain is: JEPG compression-median filter-Gaussian blur. Analyzing the tampering chain of images plays an important role in image traceability multimedia forensics, and today's various image processing software brings great convenience to the operation of digital images, and at the same time, it also significantly reduces the risk of image forgery. cost and are imperceptible to the naked eye.

分类问题很容易解决二分类问题，而对于含4种算法的篡改链而言，其排列数为

而

通过分类问题解决篡改链检测这一难题似乎更不乐观。故亟待提出一种新的图像篡改链检测方式以提高图像篡改链检测结果的准确性。The identification of existing image processing algorithms depends on the good classification performance of neural networks. Thanks to the strong representation capabilities of neural networks, the detection of tampering chains with only one image processing algorithm is basically mature. However, for the case where the image processing algorithm is greater than 2, the effect of the current classification-based algorithm is difficult to meet the application requirements. The traces on the tampered image are not obvious; (2) Order: Since the tampering chain is an ordered combination of finite algorithms, after the finite algorithms are detected, the order of the algorithms is difficult to determine. (3) Complexity: The tampering leaves intricate traces, and the classification network is difficult to implement. With the increase of the number of limited algorithms, the number of tampering chains will increase exponentially as a category of the classification network. For example, for a network with only 2 algorithms For the tampering chain, the number of permutations is

The classification problem is easy to solve the binary classification problem, and for the tampering chain with 4 algorithms, the number of permutations is

and

Solving the difficult problem of tamper chain detection through classification problems seems even less optimistic. Therefore, it is urgent to propose a new image tampering chain detection method to improve the accuracy of image tampering chain detection results.

Contents of the invention

Therefore, the technical problem to be solved by the present invention is to overcome the defect of low recognition accuracy of image tampering chains of various algorithms using neural network classification performance, thereby providing a method, device and electronic equipment for detecting image tampering chains.

According to the first aspect, the embodiment of the present invention discloses an image tampering chain detection method, comprising: inputting the image to be detected to a preset image feature extractor for feature extraction, and obtaining a target feature map corresponding to the image to be detected; The target feature map is convolved with the target number of convolution kernels, and the convolution result is flattened to obtain the word vector representation of the target dimension; the word vector representation is input to the preset image processing algorithm translation model for image processing Algorithm identification to obtain the detection result of the tampering chain of the image to be detected, wherein the translation model of the preset image processing algorithm is constructed based on the attention mechanism model.

Optionally, the preset image feature extractor includes: a first convolutional layer, a pooling layer, a residual network layer, and a second convolutional layer; the first convolutional layer is used to input the first An image to be detected with a number of channels and a first pixel size is converted into a feature map of an image to be detected with a second number of channels and a first pixel size, wherein the second number of channels is greater than the first number of channels; the pooling layer , for downsampling the feature map of the image to be detected with the second number of channels and the first pixel size to obtain the feature map of the image to be detected with the second number of channels and the second pixel size, wherein the second pixel size The number of pixels included is less than the number of pixels included in the first pixel size; the residual network layer is used to process the feature map of the image to be detected with the second channel number and the second pixel size to obtain the first An image feature map to be detected with two channels and a third pixel size, wherein the number of pixels included in the third pixel size is smaller than the number of pixels included in the second pixel size; the second convolutional layer is used for The image feature map to be detected with the second channel number and the third pixel size is down-sampled to obtain the image feature map with the third channel number and the third pixel size, and the third channel number, the third pixel size A pixel-sized feature map of the image to be detected is used as a target feature map corresponding to the image to be detected, wherein the third channel number is smaller than the second channel number and the third channel number is greater than the first channel number.

Optionally, inputting the word vector representation into a preset image processing algorithm translation model for image processing algorithm recognition includes: performing dimensionality reduction processing on the word vector representation, and inputting the dimensionality reduction processed word vector representation into the preset The image processing algorithm translation model is set up to identify the image processing algorithm.

Optionally, the number of convolution kernels is greater than the number of image processing algorithms capable of processing the image to be detected.

According to the second aspect, the embodiment of the present invention also discloses an image tampering chain detection device, including: a feature map acquisition module, which is used to input the image to be detected to a preset image feature extractor for feature extraction, and obtain the image to be detected The target feature map corresponding to the image; the word vector representation acquisition module, which is used to perform a convolution operation on the target feature map and the target number of convolution kernels, and flatten the convolution result to obtain the word vector representation of the target dimension; the detection result An acquisition module, configured to input the word vector representation into a preset image processing algorithm translation model for image processing algorithm recognition, and obtain the tampering chain detection result of the image to be detected, wherein the preset image processing algorithm translation model is based on attention The force mechanism model is constructed.

Optionally, the preset image feature extractor includes: a first convolutional layer, a pooling layer, a residual network layer, and a second convolutional layer; the first convolutional layer is used to input the first An image to be detected with a number of channels and a first pixel size is converted into a feature map of an image to be detected with a second number of channels and a first pixel size, wherein the second number of channels is greater than the first number of channels; the pooling layer , for downsampling the feature map of the image to be detected with the second number of channels and the first pixel size to obtain the feature map of the image to be detected with the second number of channels and the second pixel size, wherein the second pixel size The number of pixels included is less than the number of pixels included in the first pixel size; the residual network layer is used to process the feature map of the image to be detected with the second channel number and the second pixel size to obtain the first An image feature map to be detected with two channels and a third pixel size, wherein the number of pixels included in the third pixel size is smaller than the number of pixels included in the second pixel size; the second convolutional layer is used for The image feature map to be detected with the second channel number and the third pixel size is down-sampled to obtain the image feature map with the third channel number and the third pixel size, and the third channel number, the third pixel size A pixel-sized feature map of the image to be detected is used as a target feature map corresponding to the image to be detected, wherein the third channel number is smaller than the second channel number and the third channel number is greater than the first channel number.

Optionally, the detection result acquisition module is further configured to perform dimensionality reduction processing on the word vector representation, and input the dimensionality reduction processed word vector representation to a preset image processing algorithm translation model for image processing algorithm recognition.

Optionally, the number of convolution kernels is greater than the number of image processing algorithms capable of processing the image to be detected.

According to the third aspect, the embodiment of the present invention also discloses an electronic device, including: at least one processor; and a memory connected in communication with the at least one processor; wherein, the memory stores information that can be used by the at least one processor An instruction executed by a processor, the instruction is executed by the at least one processor, so that the at least one processor executes the image tampering chain detection method described in the first aspect or any optional implementation manner of the first aspect step.

According to the fourth aspect, the embodiment of the present invention also discloses a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the first aspect or any optional implementation of the first aspect can be realized. The steps of the image tampering chain detection method described in the manner.

The technical solution of the present invention has the following advantages:

The image tampering chain detection method/device provided by the present invention, by inputting the image to be detected into the preset image feature extractor for feature extraction, obtains the target feature map corresponding to the image to be detected, and combines the target feature map with the number of target convolution kernels Perform a convolution operation, flatten the convolution result to obtain the word vector representation of the target dimension, input the word vector representation into the preset image processing algorithm translation model based on the attention mechanism model, and perform image processing algorithm recognition to obtain the target dimension The tampering chain detection result of the image; compared with using the classification performance of the neural network to directly classify and detect the image tampering chain, the method provided by the present invention extracts features on the image to be detected, and performs convolution on the extracted target feature map Product and flattening operations, the word vector representation obtained after flattening is input to the translation model of the preset image processing algorithm so that the translation model of the preset image processing algorithm can directly translate the image processing algorithm corresponding to the image to be detected. The output of the translation results makes the number of possible tampering chains corresponding to the image to be detected decrease exponentially, which reduces the complexity of tampering chain detection under multi-algorithms. For tampering chain detection tasks, the translation model is more in line with human recognition. Knowing, stripping away cocoons, translating word by word, under the task of detecting image tampering chains of the same length, compared with classification models, using translation models to detect tampering chains has higher accuracy. At the same time, the preset image processing algorithm translation model is based on The attention mechanism model is constructed, which makes the ordering result of the tampering chain translated based on the preset image processing algorithm translation model more accurate.

Description of drawings

In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the accompanying drawings that need to be used in the specific implementation or description of the prior art. Obviously, the accompanying drawings in the following description The drawings show some implementations of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative work.

Fig. 1 is the flow chart of a specific example of image tampering chain detection method in the embodiment of the present invention;

Fig. 2 is a schematic diagram of a specific network architecture of an image tampering chain detection method in an embodiment of the present invention;

Fig. 3 is a schematic diagram of a specific image feature extractor of the image tampering chain detection method in the embodiment of the present invention;

4 is a schematic diagram of a specific residual network layer of the image tampering chain detection method in the embodiment of the present invention;

FIG. 5 is a functional block diagram of a specific example of an image tampering chain detection device in an embodiment of the present invention;

Fig. 6 is a diagram of a specific example of an electronic device in an embodiment of the present invention.

Detailed ways

The technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Apparently, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" etc. The indicated orientation or positional relationship is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, or in a specific orientation. construction and operation, therefore, should not be construed as limiting the invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and should not be construed as indicating or implying relative importance.

In the description of the present invention, it should be noted that unless otherwise specified and limited, the terms "installation", "connection" and "connection" should be understood in a broad sense, for example, it can be a fixed connection or a detachable connection. Connected, or integrally connected; it can be mechanically or electrically connected; it can be directly connected, or indirectly connected through an intermediary, or it can be the internal communication of two components, which can be wireless or wired connect. Those of ordinary skill in the art can understand the specific meanings of the above terms in the present invention in specific situations.

In addition, the technical features involved in the different embodiments of the present invention described below may be combined with each other as long as there is no conflict with each other.

The embodiment of the present invention discloses an image tampering chain detection method, as shown in Figure 1 and Figure 2, the method includes the following steps:

Step 101, input the image to be detected to a preset image feature extractor for feature extraction, and obtain a target feature map corresponding to the image to be detected.

Exemplarily, the image to be detected may be an image not processed by any image processing algorithm, or an image processed by at least one image processing algorithm, wherein the image processing algorithm may include but not limited to JPEG compression, median filter , Gaussian blur, image super-resolution, adding noise, etc. The obtained image to be detected is input to the preset image feature extractor for feature extraction, and the target feature map corresponding to the image to be detected is obtained. The embodiment of the present application does not limit the specific structure of the preset image feature extractor. Those skilled in the art It can be determined according to actual needs, for example, a corresponding image feature extractor can be selected according to the size of the target feature map to be obtained. In the embodiment of the present application, a 256×256×1 image to be detected is selected, and the size of the target feature map is 32×32×32.

Step 102, performing a convolution operation on the target feature map and a target number of convolution kernels, and flattening the convolution result to obtain a word vector representation of the target dimension.

Exemplarily, each channel of the extracted target feature map hides the traces of various image processing algorithms, so the target feature map is convolved with the target number of convolution kernels, and the result obtained by the convolution is unfolded The word vector representation of C×H*W is obtained by the flattening process, and the word vector representation obtained after the flattening process can also be called a “sentence vector”.

As an optional implementation manner of the present invention, the number of the convolution kernels is greater than the number of image processing algorithms capable of processing the image to be detected.

Step 103, input the word vector representation into the preset image processing algorithm translation model for image processing algorithm recognition, and obtain the tampering chain detection result of the image to be detected, wherein the preset image processing algorithm translation model is based on the attention mechanism The model is constructed.

Exemplarily, the obtained word vector representation containing tampering traces is input to the translation model of the preset image processing algorithm for encoding and decoding processing, and the tampering chain detection result of the image to be detected is obtained. The embodiment of the present application does not limit the translation model of the preset image processing algorithm. Those skilled in the art can construct the translation model based on the attention mechanism model according to actual usage requirements, or choose a Transformer network. The preset image processing algorithm translation model can be obtained by pre-training according to the category of the image processing algorithm that may correspond to the image to be translated. During the model training process, prior information (Dec _in ) can be used to make the Transformer translate The model can realize parallel training and improve the training efficiency of the model.

For example, there are five types of image processing algorithms that may correspond to the image to be translated: JPEG compression, median filter, Gaussian blur, super-resolution, and adding Gaussian noise, respectively represented by JP, MF, GB, SR, and GN , build the corresponding dictionary as shown in Table 1, if an image has undergone the above five operations in sequence, its tampering chain is JP-MF-GB-SR-GN, and the corresponding real label y=[0, 1,2,3,4].

Table 1 Processing Algorithm Dictionary

The loss function used in the embodiment of this application is shown in the following formula:

逼近真实篡改链y。The image processing algorithm translation model trained by optimizing the cross-entropy loss function makes the predicted tampering chain

Approximate the real tampering chain y.

Compared with using the classification performance of the neural network to classify and detect image tampering chains, the method provided by the embodiment of the present invention extracts features directly on the image to be detected, and performs convolution and flattening operations on the extracted target feature map , the word vector representation obtained after flattening is input to the translation model of the preset image processing algorithm so that the translation model of the preset image processing algorithm directly translates the image processing algorithm corresponding to the image to be detected, and the output of the translation result makes The number of possible tampering chains corresponding to the image to be detected decreases exponentially, which reduces the complexity of tampering chain detection under multi-algorithms, and the translation model of the preset image processing algorithm is constructed based on the attention mechanism model, so that based on the The sequence result of the tampering chain translated by the preset image processing algorithm translation model is more accurate.

Aiming at the order and complexity of tampering chains, the image tampering chain detection method based on machine translation technology provided by the embodiment of the present invention is no longer limited to the classification problem of neural networks to detect tampering chains of tampered images, but Combining the advantages of machine translation, the limited algorithm contained in the tampered image is translated in an orderly manner one by one "word" to obtain its corresponding tampering chain, which conforms to people's cognition of handling affairs, breaks the inherent thinking of the original classification algorithm, and can Efficiently detect tampering chains corresponding to tampered images; with the progress of machine translation, the number of tampering chains will also decrease exponentially due to the translation of the previous "words", which solves the order and complexity of tampering chain storage Difficulties, for example, for a tampering chain containing five image processing algorithms, when the first "word" is decoded, the possible decoding results are reduced from the original 120 to 24, and the translation results are quite impressive. Improved image tampering chain detection efficiency.

As an optional implementation of the present invention, the preset image feature extractor includes: a first convolutional layer, a pooling layer, a residual network layer, and a second convolutional layer; The number of convolutional layers, pooling layers, and residual network layers included in the feature extractor is not limited, and can be determined by those skilled in the art according to actual needs. The structure of the image feature extractor used in the embodiment of the present application is shown in FIG. 3 .

The first convolutional layer is used to convert the input image to be detected with the first number of channels and the first pixel size into a feature map of the image to be detected with the second number of channels and the first pixel size, wherein the second channel The number is greater than the first channel number;

The pooling layer is used to perform downsampling processing on the feature map of the image to be detected with the second number of channels and the first pixel size to obtain the feature map of the image to be detected with the second number of channels and the second pixel size, wherein the The number of pixels included in the second pixel size is smaller than the number of pixels included in the first pixel size;

The residual network layer is used to process the feature map of the image to be detected with the second number of channels and the second pixel size to obtain the feature map of the image to be detected with the second number of channels and the third pixel size, wherein the the number of pixels contained in the third pixel size is less than the number of pixels contained in the second pixel size;

The second convolutional layer is configured to perform downsampling processing on the feature map of the image to be detected with the second number of channels and the third pixel size to obtain the feature map of the image to be detected with the third number of channels and the third pixel size, Using the feature map of the image to be detected with the third channel number and the third pixel size as the target feature map corresponding to the image to be detected, wherein the third channel number is smaller than the second channel number and the third channel number greater than the first channel number.

Exemplarily, the embodiment of the present application does not limit the pixel size and the number of channels of the image feature map of each network layer above, which can be determined by those skilled in the art according to actual needs. Specifically, as shown in Figure 3, the input of the image feature extractor is a 256×256×1 image, and the number of channels is 1, and the H×W×C is 32×32×32 obtained by extracting features through the image feature extractor. feature map of . In the image feature extractor, the first convolutional layer (Conv) converts the input image (256×256×1) from 1 channel to 64 channels (256×256×64), and then passes through the maximum pooling layer (MaxPool ) downsampled output to obtain (128×128×64), then output through the residual network layer to obtain (32×32×64), and finally pass through a layer of convolutional layer for downsampling output feature map (32×32×32). In the embodiment of the present application, the residual network layer (ResNet) is composed of 16 residual blocks and 2 maximum pooling layers, and the specific structure is shown in FIG. 4 . A deeper convolutional neural network composed of a residual network can better extract the features left by the image processing algorithm and construct a more suitable sentence vector.

As an optional embodiment of the present invention, inputting the word vector representation into a preset image processing algorithm translation model for image processing algorithm recognition includes: performing dimensionality reduction processing on the word vector representation, and converting the dimensionality reduction processed word The vector representation is input to the preset image processing algorithm translation model for image processing algorithm identification. In order to improve the efficiency of machine translation, dimensionality reduction is performed on the word vector representation input to the translation model of the preset image processing algorithm to reduce redundant feature vectors in the word vector representation, so as to ensure the accuracy of translation results and improve translation efficiency.

As a specific implementation of the present application, the original image is divided into several 256×256 sub-images, and then they are tampered in a certain order to obtain the corresponding tampering chain as prior information. It is assumed that the tampering algorithm includes Gaussian blur, median filter, super-resolution, JPEG compression, and Gaussian noise (GB-MF-RS-JP-GN), where Gaussian blur and median filter use a 3×3 template for filtering; super-resolution uses The algorithm is to reduce the size of the image to a quarter of the original size first, and then enlarge it four times through the bilinear difference algorithm for super-resolution; the compression factor QF of JPEG compression is 70; the added Gaussian noise is 0 mean , with a variance in the range of 5-10. The accuracy translation results of the image tampering chain detection algorithm provided by the above embodiments are shown in Table 2 below:

Table 2 The accuracy of image tampering chain prediction

Among them, A is the number of permutations; Accuracy is the accuracy rate of tampering chains that are completely consistent; Bilingual Evaluation Replacement Score (BLEU), the score of a perfect match is 1.0, and the score of a complete mismatch is 0.0.

It can be seen from Table 2 that due to the concealment of tampering traces, with the increase of the total number of algorithms, the tampering traces left by the previous algorithm are covered by the later algorithm, resulting in a decrease in accuracy, but as can be seen from the BLEU index, its performance index is still considerable . The following table 3 is the confusion matrix of the detection results of the tampering chain based on the previous optimal classification algorithm MISLnet, and the following table 4 is the confusion matrix of the detection results of the tampering chain based on the technical solution of this application:

Table 3 shows the detection results of the tampering chain based on the previous optimal classification algorithm MISLnet

0R MF GB RS MF-GB GB-MF MF-RS RS-MF GB-RS RS-GB 0R 99.16 0.42 0.00 0.39 0.00 0.03 0.00 0.00 0.00 0.00 MF 0.09 88.40 0.00 0.00 0.00 0.20 0.66 10.37 0.29 0.00 GB 0.00 0.00 89.76 0.18 0.44 0.00 0.21 0.00 0.00 9.42 RS 0.03 0.00 0.55 99.25 0.00 0.00 0.14 0.03 0.00 0.00 MF-GB 0.00 0.00 0.11 0.00 94.21 1.67 0.58 0.00 2.64 0.78 GB-MF 0.00 0.03 0.00 0.00 0.60 91.41 1.75 5.49 0.72 0.00 MF-RS 0.00 0.00 0.00 0.14 0.00 0.03 99.80 0.00 0.03 0.00 RS-MF 0.00 2.97 0.06 0.06 0.00 16.28 2.17 76.01 2.45 0.00 GB-RS 0.00 0.00 0.00 0.00 0.15 0.09 0.15 0.00 99.44 0.18 RS-GB 0.00 0.00 2.95 0.00 0.09 0.00 0.00 0.00 0.79 96.17

Table 4 is the detection result of tampering chain based on the technical solution of this application

0R MF GB RS MF-GB GB-MF MF-RS RS-MF GB-RS RS-GB 0R 100.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 MF 0.00 96.06 0.00 0.00 0.00 0.16 0.13 3.65 0.00 0.00 GB 0.00 0.00 94.97 0.17 0.10 0.00 0.00 0.00 0.00 4.76 RS 0.00 0.00 0.19 99.68 0.00 0.00 0.13 0.00 0.00 0.00 MF-GB 0.00 0.00 0.03 0.00 98.69 1.25 0.00 0.00 0.00 0.03 GB-MF 0.00 0.13 0.00 0.00 0.58 96.96 0.00 2.23 0.10 0.00 MF-RS 0.00 0.00 0.00 0.10 0.00 0.00 99.90 0.00 0.00 0.00 RS-MF 0.00 3.36 0.00 0.00 0.00 4.65 0.13 91.72 0.13 0.00 GB-RS 0.00 0.00 0.00 0.00 0.00 0.13 0.00 0.03 99.78 0.06 RS-GB 0.00 0.00 2.68 0.00 0.03 0.00 0.00 0.03 0.32 96.94

Combining the data recorded in the above Table 3 and Table 4, it can be seen that the scheme recorded in the embodiment of the present application has a higher accuracy of the tampering chain detection results; and in combination with Table 2-Table 4, it can be seen that for short chains, The performance of the solution described in the embodiment of the present application is obviously better than that of the MISLnet network, and the detection results for longer chains are also very impressive.

The method provided in the embodiment of the present application can be implemented using the deep learning framework Pytorch, or can be implemented using deep learning frameworks such as TensorFlow and Caffe. Given an image to be detected, the method provided by the embodiment of this application can detect all the image processing algorithms and their sequences that it has experienced, breaking the inherent thinking of tampering chain detection, and proposing to use machine translation technology to process the tampering chain one by one. "Word" translation is of great significance to the research of image forgery detection, image traceability analysis, and multimedia security.

The embodiment of the present invention also discloses an image tampering chain detection device, as shown in Figure 5, the device includes:

A feature map acquisition module 501, configured to input the image to be detected to a preset image feature extractor for feature extraction, to obtain a target feature map corresponding to the image to be detected;

A word vector representation acquisition module 502, configured to perform a convolution operation on the target feature map and a target number of convolution kernels, and flatten the convolution result to obtain a word vector representation of the target dimension;

The detection result acquisition module 503 is used to input the word vector representation into the translation model of the preset image processing algorithm to identify the image processing algorithm, and obtain the tampering chain detection result of the image to be detected, wherein the preset image processing algorithm translates The model is constructed based on the attention mechanism model.

The image tampering chain detection device provided by the present invention, compared to using the classification performance of the neural network to classify and detect the image tampering chain, the device provided by the embodiment of the present invention directly extracts features on the image to be detected, and extracts the extracted The target feature map is convolved and flattened, and the word vector representation obtained after the flattening is input to the preset image processing algorithm translation model so that the preset image processing algorithm translation model directly corresponds to the image processing algorithm corresponding to the image to be detected. Translate, with the output of the translation results, the number of possible tampering chains corresponding to the image to be detected decreases exponentially, reducing the complexity of tampering chain detection under multi-algorithms, and the preset image processing algorithm translation model is based on The attention mechanism model is constructed, which makes the ordering result of the tampering chain translated by the translation model based on the preset image processing algorithm more accurate.

As an optional implementation of the present invention, the preset image feature extractor includes: a first convolutional layer, a pooling layer, a residual network layer, and a second convolutional layer;

The first convolutional layer is used to convert the input image to be detected with the first number of channels and the first pixel size into a feature map of the image to be detected with the second number of channels and the first pixel size, wherein the second channel The number is greater than the first channel number;

The pooling layer is used to perform downsampling processing on the feature map of the image to be detected with the second number of channels and the first pixel size to obtain the feature map of the image to be detected with the second number of channels and the second pixel size, wherein the The number of pixels included in the second pixel size is smaller than the number of pixels included in the first pixel size;

The residual network layer is used to process the feature map of the image to be detected with the second number of channels and the second pixel size to obtain the feature map of the image to be detected with the second number of channels and the third pixel size, wherein the the number of pixels contained in the third pixel size is less than the number of pixels contained in the second pixel size;

The second convolutional layer is configured to perform downsampling processing on the feature map of the image to be detected with the second number of channels and the third pixel size to obtain the feature map of the image to be detected with the third number of channels and the third pixel size, Using the feature map of the image to be detected with the third channel number and the third pixel size as the target feature map corresponding to the image to be detected, wherein the third channel number is smaller than the second channel number and the third channel number greater than the first channel number.

As an optional embodiment of the present invention, the detection result acquisition module is also used to perform dimensionality reduction processing on the word vector representation, and input the dimensionality reduction processed word vector representation into a preset image processing algorithm translation model for image processing. Process Algorithm Identification.

As an optional implementation manner of the present invention, the number of the convolution kernels is greater than the number of image processing algorithms capable of processing the image to be detected.

The embodiment of the present invention also provides an electronic device. As shown in FIG. 6, the electronic device may include a processor 601 and a memory 602, wherein the processor 601 and the memory 602 may be connected through a bus or in other ways. In FIG. Take the bus connection as an example.

The processor 601 may be a central processing unit (Central Processing Unit, CPU). The processor 601 can also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application-specific integrated circuits (Application Specific Integrated Circuit, ASIC), field-programmable gate array (Field-Programmable Gate Array, FPGA) or Other chips such as programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above-mentioned types of chips.

The memory 602, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions/programs corresponding to the image tampering chain detection method in the embodiment of the present invention. module. The processor 601 executes various functional applications and data processing of the processor by running non-transitory software programs, instructions and modules stored in the memory 602, that is, implements the image tampering chain detection method in the above method embodiments.

The memory 602 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application program required by at least one function; the data storage area may store data created by the processor 601 and the like. In addition, the memory 602 may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory 602 may optionally include memory that is remotely located relative to the processor 601, and these remote memories may be connected to the processor 601 through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The one or more modules are stored in the memory 602, and when executed by the processor 601, the image tampering chain detection method in the embodiment shown in FIG. 1 is executed.

Specific details of the above-mentioned electronic device can be understood by referring to corresponding descriptions and effects in the embodiment shown in FIG. 1 , and details are not repeated here.

Those skilled in the art can understand that all or part of the processes in the methods of the above-mentioned embodiments can be completed by instructing related hardware through computer programs, and the programs can be stored in a computer-readable storage medium. During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium can be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a flash memory (Flash Memory), a hard disk (Hard Disk Drive) , abbreviation: HDD) or solid-state hard disk (Solid-State Drive, SSD) etc.; The storage medium may also include a combination of the above-mentioned types of memory.

Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations all fall into the scope of the appended claims. within the limited range.

Claims (8)

1. An image tamper chain detection method, comprising:

inputting an image to be detected into a preset image feature extractor for feature extraction to obtain a target feature map corresponding to the image to be detected;

performing convolution operation on the target characteristic graph and convolution kernels of the target number, and flattening the convolution result to obtain word vector representation of the target dimension;

inputting the word vector representation into a preset image processing algorithm translation model for image processing algorithm recognition to obtain a tampering chain detection result of the image to be detected, wherein the preset image processing algorithm translation model is constructed on the basis of an attention mechanism model; the preset image feature extractor includes: a first convolution layer, a pooling layer, a residual network layer and a second convolution layer;

the first convolution layer is used for converting an input first channel number and an input image to be detected with a first pixel size into a second channel number and an input image feature map to be detected with the first pixel size, wherein the second channel number is larger than the first channel number;

the pooling layer is used for performing down-sampling processing on the second channel number and the to-be-detected image feature map with the first pixel size to obtain a second channel number and a to-be-detected image feature map with the second pixel size, wherein the number of pixels contained in the second pixel size is smaller than that contained in the first pixel size;

the residual error network layer is used for processing the second channel number and the to-be-detected image feature map with the second pixel size to obtain a to-be-detected image feature map with the second channel number and the third pixel size, wherein the number of pixels contained in the third pixel size is smaller than that contained in the second pixel size;

the second convolutional layer is used for performing downsampling processing on the second channel number and the image feature map to be detected with the third pixel size to obtain a third channel number and a third pixel size image feature map to be detected, and the third channel number and the image feature map to be detected with the third pixel size are used as target feature maps corresponding to the image to be detected, wherein the third channel number is smaller than the second channel number and the third channel number is larger than the first channel number.

2. The method of claim 1, wherein inputting the word vector representation into a pre-defined image processing algorithm translation model for image processing algorithm recognition comprises:

and performing dimension reduction on the word vector representation, and inputting the word vector representation subjected to dimension reduction into a preset image processing algorithm translation model for image processing algorithm identification.

3. Method according to claim 1 or 2, characterized in that the number of convolution kernels is greater than the number of image processing algorithms capable of processing the image to be detected.

4. An image tamper chain detection device, comprising:

the characteristic diagram acquisition module is used for inputting an image to be detected into a preset image characteristic extractor for characteristic extraction to obtain a target characteristic diagram corresponding to the image to be detected;

the word vector representation acquisition module is used for performing convolution operation on the target characteristic graph and the convolution kernels of the target quantity, and flattening the convolution result to obtain word vector representation of the target dimension;

the detection result acquisition module is used for inputting the word vector representation into a preset image processing algorithm translation model for image processing algorithm recognition to obtain a tampering chain detection result of the image to be detected, wherein the preset image processing algorithm translation model is constructed on the basis of an attention mechanism model; the preset image feature extractor includes: a first convolution layer, a pooling layer, a residual network layer and a second convolution layer;

the first convolution layer is used for converting an input first channel number and an input image to be detected with a first pixel size into a second channel number and an input image feature map to be detected with the first pixel size, wherein the second channel number is larger than the first channel number;

the pooling layer is used for performing down-sampling processing on the second channel number and the to-be-detected image feature map with the first pixel size to obtain a second channel number and a to-be-detected image feature map with a second pixel size, wherein the number of pixels contained in the second pixel size is smaller than that contained in the first pixel size;

the residual error network layer is used for processing the second channel number and the image characteristic diagram to be detected with the second pixel size to obtain a second channel number and an image characteristic diagram to be detected with a third pixel size, wherein the number of pixels contained in the third pixel size is smaller than the number of pixels contained in the second pixel size;

the second convolutional layer is used for performing downsampling processing on the second channel number and the image feature map to be detected with the third pixel size to obtain a third channel number and a third pixel size image feature map to be detected, and the third channel number and the image feature map to be detected with the third pixel size are used as target feature maps corresponding to the image to be detected, wherein the third channel number is smaller than the second channel number and the third channel number is larger than the first channel number.

5. The apparatus according to claim 4, wherein the detection result obtaining module is further configured to perform dimension reduction processing on the word vector representation, and input the word vector representation after dimension reduction processing into a preset image processing algorithm translation model for image processing algorithm recognition.

6. The apparatus according to claim 4 or 5, characterized in that the number of convolution kernels is larger than the number of image processing algorithms capable of processing the image to be detected.

7. An electronic device, comprising: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to cause the at least one processor to perform the steps of the image tamper chain detection method according to any one of claims 1-3.

8. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the image tamper chain detection method according to any one of claims 1 to 3.

Images