Image modification area positioning method, system and storage medium based on deep learning

Abstract

The present invention discloses a method, system and storage medium for locating an image modification region based on deep learning, which comprises extracting an image residual of a target RGB image, extracting the inconsistency feature of the high frequency component between the modified region and the unmodified region in the target RGB image based on the image residual using a first neural network, and generating a first feature map according to the extracted inconsistency feature; extracting the local binary pattern feature of the target RGB image, merging the local binary pattern feature with the target RGB image to obtain a merged image, extracting the block similarity feature between the modified region and the unmodified region in the merged image using a second neural network, and generating a second feature map according to the extracted block similarity feature; fusing the first feature map and the second feature map to obtain a target feature map, and locating the image modification region of the target RGB image based on the target feature map. The method can efficiently and accurately locate the modification region in the image.

Description

Technical Field

[0001] The present invention relates to the field of computer image processing technology, and in particular to a method, system and storage medium for locating an image modification area based on deep learning. Background Art

[0002] Image restoration refers to reconstructing the lost or damaged parts of an image, using known information of the damaged image and filling the damaged area according to certain rules, so that the restored image can achieve or approach the visual effect of the original image. Image restoration technology can be used to modify an image, but it is difficult to visually see that the modified image has been modified, which affects people's judgment of the authenticity of the image.

[0003] There are many traditional methods for locating image modification areas. For example, Wu Q, Sun SJ, Zhu W, et al. Detection of digital doctoring in exemplar-based inpainted images [C] / / 2008 International Conference on Machine Learning and Cybernetics. IEEE, 2008, 3: 1222-1226. proposed a blind detection method based on zero connectivity features and fuzzy membership to detect specific tampering modifications. First, the block pairs are marked with zero connectivity to obtain the matching features of all blocks in the suspicious area, and the fuzzy membership is calculated by constructing a semi-trapezoidal membership function. Then, the tampered area is identified by cut set partitioning. Bacchuwar KS, Ramakrishnan K RA jump patch-block matching algorithm for multiple forgery detection[C] / / 2013 International Mutli-Conference on A utomation, Computing, Communication, Control and Compressed Sensing(iMac4s).IEEE,2013:723-728. proposed a blind detection method based on the comparison of the median of the block in the brightness component of the image and the suspicious region (Region of Suspicion, ROS). The above traditional method of positioning sample block modification requires a lot of time for calculation and even requires manual operation.

[0004] The method of locating sample block modification based on deep learning can make up for the above problems. For example, Zhu X, Qian Y, Zhao X, et al. A deep learning approach to patch-based image inpainting forensics [J]. Signal Processing: Image Communication, 2018, 67: 90-99. proposed a method of using convolutional neural network (CNN) to detect sample block image modification. The convolutional neural network is constructed according to the encoder-decoder network structure and can predict the modification probability of each pixel in the image. Through the established convolutional neural network, when collecting evidence for the modified area, there is no need to consider feature extraction and classifier design, nor is there any need for post-processing like traditional forensics methods. Although the above method can greatly reduce the detection time while maintaining accuracy, the existing deep learning method for locating sample block modification is not mature enough, and the location of the modified area boundary is not clear enough when detecting real modified images. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention provides an image modification area positioning method, system and storage medium based on deep learning, which can efficiently and accurately locate the modification area in the image.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0007] In a first aspect, the present invention provides an image modification region positioning method based on deep learning, comprising the following steps:

[0008] Acquire a target RGB image, where the target RGB image is a modified RGB image;

[0009] Extracting an image residual of the target RGB image;

[0010] Based on the image residual, using a first neural network to extract inconsistency features of high-frequency components between a modified area and an unmodified area in the target RGB image, and generating a first feature map according to the extracted inconsistency features;

[0011] Extracting local binary pattern features of the target RGB image;

[0012] Merging the local binary pattern feature with the target RGB image to obtain a merged image, extracting block similarity features between the modified area and the unmodified area in the merged image using a second neural network, and generating a second feature map according to the extracted block similarity features;

[0013] The first feature map and the second feature map are fused to obtain a target feature map, and an image modification region of the target RGB image is located based on the target feature map.

[0014] Further, the step of extracting block similarity features between the modified region and the unmodified region in the merged image using a second neural network comprises the steps of:

[0015] Performing a convolution operation on the merged image through two consecutive convolution blocks and three dual attention convolution blocks to obtain three feature maps;

[0016] The three feature maps are early fused to extract block similarity features between the modified region and the unmodified region in the merged image.

[0017] Further, the extracting of local binary pattern features of the target RGB image comprises the steps of:

[0018] The local binary pattern LBP is used to extract the local binary pattern features of the target RGB image, wherein the calculation formula of the local binary pattern LBP is:

[0019]

[0020] Among them, (x c ,y c ) represents the center pixel, i c Represents the gray value of the center pixel, i p Represents the grayscale value of the neighboring pixels of the central pixel.

[0021] Further, generating a second feature map according to the extracted block similarity features comprises the steps of:

[0022] A transposed convolution with two times upsampling and a transposed convolution with four times upsampling are used to enlarge the size of the second feature map to the size of the target RGB image; and a convolution layer with a stride of 1 and a size of 5×5 is used at the same time to reduce the checkerboard artifacts caused by the transposed convolution.

[0023] Further, the extracting the image residual of the target RGB image comprises the steps of:

[0024] Using three first-order high-pass filters with initialized kernels to perform convolution operations on the three channels of the target RGB image respectively, to obtain three convolution results;

[0025] The three convolution results are connected by depth to obtain the residual image.

[0026] Further, the extracting the inconsistency feature of the high-frequency component between the modified area and the unmodified area in the target RGB image using the first neural network comprises the steps of:

[0027] The inconsistency features of high-frequency components between the modified area and the unmodified area in the target RGB image are extracted through four ResNet blocks, wherein each of the ResNet blocks includes two bottleneck units, and each of the bottleneck units includes three consecutive convolutional layers and an identity shortcut connection.

[0028] Further, generating a first feature map according to the extracted inconsistency features comprises the steps of:

[0029] The size of the first feature map is enlarged to the size of the target RGB image through two quadruple upsampling transposed convolutions, and a convolution layer with a stride of 1 and a size of 5×5 is used to reduce the checkerboard artifacts caused by the transposed convolution.

[0030] Further, the fusing of the first feature map and the second feature map to obtain a target feature map, and locating the image modification area of ​​the target RGB image based on the target feature map, comprises the steps of:

[0031] The first feature map and the second feature map are concatenated according to the depth to obtain a fused feature map;

[0032] The fused feature map is respectively subjected to convolutional filters of sizes 1×1, 3×3, and 5×5 to obtain three feature maps of different sizes;

[0033] The three feature maps of different sizes are spliced ​​according to depth to obtain the target feature map;

[0034] Logistic regression is used to perform pixel-by-pixel prediction on the target feature map to locate the image modification area of ​​the target RGB image.

[0035] In a second aspect, the present invention provides an image modification region positioning system based on deep learning, comprising:

[0036] An image acquisition unit, used for acquiring a target RGB image, wherein the target RGB image is a modified RGB image;

[0037] A first pre-filtering unit, used to extract an image residual of the target RGB image;

[0038] A first feature extraction unit is used to extract inconsistency features of high-frequency components between a modified area and an unmodified area in the target RGB image based on the image residual using a first neural network, and generate a first feature map according to the extracted inconsistency features;

[0039] A second pre-filtering unit, used for extracting local binary pattern features of the target RGB image;

[0040] A second feature extraction unit is used to merge the local binary pattern feature with the target RGB image to obtain a merged image, extract block similarity features between the modified area and the unmodified area in the merged image using a second neural network, and generate a second feature map according to the extracted block similarity features;

[0041] The feature fusion unit is used to fuse the first feature map and the second feature map to obtain a target feature map, and locate the image modification area of ​​the target RGB image based on the target feature map.

[0042] In a third aspect, the present invention provides a computer-readable storage medium, which stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the image modification area positioning method based on deep learning as described above.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] In the first aspect of the present application, a method for locating an image modification region based on deep learning is proposed. The method extracts the image residual of a target RGB image, and uses a first neural network to extract the inconsistency features of the high-frequency components between the modified region and the unmodified region in the target RGB image, and generates a first feature map based on the extracted inconsistency features; extracts the local binary pattern features of the target RGB image, merges the local binary pattern features with the target RGB image to obtain a merged image, and uses a second neural network to extract the block similarity features between the modified region and the unmodified region in the merged image, and generates a second feature map based on the extracted block similarity features, and finally fuses the first feature map and the second feature map. The method not only extracts modification traces based on the inconsistency of the high-frequency components of the target RGB image, but also extracts modification traces based on the block similarity in the texture features of the target RGB image. Therefore, the method can efficiently and accurately locate the modified region in the image.

[0045] It can be understood that the beneficial effects of the second to third aspects compared with the related art are the same as the beneficial effects of the first aspect compared with the related art. Please refer to the relevant description in the first aspect, and no further details will be given here. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0047] Figure 1 A flowchart of a method for locating an image modification area based on deep learning provided by an embodiment of the present invention;

[0048] Figure 2 A flowchart of obtaining a second feature map provided by an embodiment of the present invention;

[0049] Figure 3 A structural diagram of dual attention provided for one embodiment of the present invention;

[0050] Figure 4 A structural diagram of an image modification region positioning system based on deep learning provided for one embodiment of the present invention. DETAILED DESCRIPTION

[0051] The technical solutions of the embodiments of the present disclosure will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure. It should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other in the absence of conflict. In addition, the purpose of the drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present disclosure, but it cannot be understood as a limitation on the scope of protection of the present disclosure.

[0052] In the description of the present invention, features defined as "first" or "second" may include one or more of the features explicitly or implicitly. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0053] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0054] Traditional methods of locating image modification areas require a lot of time for calculation and even require manual operations. Deep learning methods can greatly reduce detection time while maintaining accuracy, but existing deep learning methods are not mature enough and the positioning of the modified area boundaries is not clear enough when detecting real modified images.

[0055] In order to solve the above problems, the present application not only extracts the image residual of the target RGB image, but also uses a first neural network to extract the inconsistency features of the high-frequency components between the modified area and the unmodified area in the target RGB image based on the image residual; it also extracts the local binary pattern features of the target RGB image, merges the local binary pattern features with the target RGB image to obtain a merged image, and uses a second neural network to extract the block similarity features between the modified area and the unmodified area in the merged image. A first feature map is generated based on the extracted inconsistency features, a second feature map is generated based on the extracted block similarity features, and finally the first feature map and the second feature map are merged. In summary, the present application not only extracts the modified area based on the inconsistency of the high-frequency components of the target RGB image, but also extracts the modified area based on the block similarity in the texture features of the target RGB image. Therefore, the present application can efficiently and accurately locate the modified area in the image.

[0056] Reference Figures 1 to 3 This embodiment provides a method for locating image modification traces based on deep learning, and the specific implementation steps are as follows:

[0057] Step S100: Acquire a target RGB image, where the target RGB image is a modified RGB image.

[0058] Specifically, the RGB image is a modified image having a modified area.

[0059] Step S200: extracting the image residual of the target RGB image.

[0060] Specifically, the filter kernels of three first-order high-pass filters are initialized, and the three channels of the target RGB image are convolved using the three first-order high-pass filters with the initialized kernels to obtain three convolution results and concatenate the three convolution results by depth to obtain a residual image. The high-pass filter is set as a learnable filter through parameter setting during the training process, and the parameter is adjusted using gradient descent so that the high-pass filter can automatically obtain the optimal weight parameter. The filter kernels of the three first-order high-pass filters are initialized as follows:

[0061]

[0062] Step S300: Based on the image residual, use the first neural network to extract the inconsistency features of the high-frequency components between the modified area and the unmodified area in the target RGB image, and generate a first feature map according to the extracted inconsistency features.

[0063] Specifically, in order to better extract the modification traces in the target RGB image, the target RGB image is subjected to high-pass filtering to obtain a residual image with only high-frequency components. Based on the residual image, a first neural network including four ResNet blocks is used to extract the inconsistency features of the high-frequency components between the modified area and the unmodified area in the target RGB image, and a first feature map is generated according to the extracted inconsistency features. Referring to Table 1, each ResNet block is provided with two bottleneck units. Each bottleneck unit is provided with three consecutive convolutional layers and an identical shortcut connection. The sizes of the three consecutive convolutional layers are 1×1, 3×3, and 1×1, respectively. The three consecutive convolutional layers perform batch normalization and ReLu activation before performing the convolution operation. The stride of the first five convolutional layers in each ResNet block is set to 1, and the stride of the last convolutional layer is set to 2.

[0064]

[0065] Table 1

[0066] Among them, the size of the first feature map is enlarged to the target RGB image size by performing two four-fold upsampling transposed convolutions, and a convolution layer with a stride of 1 and a size of 5×5 is used to reduce the checkerboard artifacts caused by the transposed convolution.

[0067] Step S400: extracting local binary pattern features of the target RGB image.

[0068] Specifically, the local binary pattern features of the target RGB image are extracted by the local binary pattern LBP, wherein the formula for the local binary pattern LBP operation is:

[0069]

[0070] Among them, (x c ,y c ) represents the center pixel, i c Represents the gray value of the center pixel, i p Represents the grayscale value of the neighboring pixels of the central pixel.

[0071] Step S500: merge the local binary pattern features with the target RGB image to obtain a merged image, use a second neural network to extract block similarity features between the modified area and the unmodified area in the merged image, and generate a second feature map based on the extracted block similarity features.

[0072] Specifically, refer to Figure 2, merge the local binary pattern features with the target RGB image to obtain a merged image, use a second neural network including two consecutive convolution blocks and three dual attention convolution blocks to extract block similarity features between the modified area and the unmodified area in the merged image, and generate a second feature map according to the extracted block similarity features.

[0073] Among them, the merged image is convolved by two consecutive convolution blocks, and then three dual attention convolution blocks are used to obtain three feature maps. These three feature maps are early fused to extract the block similarity features between the modified area and the unmodified area in the merged image. Using early fusion can better extract image modification traces.

[0074] As shown in Table 2, the two convolution blocks perform ReLu activation after convolution operation, and each convolution block uses maximum pooling to reduce parameters. Only the last dual attention block of the three dual attention blocks performs maximum pooling. Using maximum pooling can avoid the reduction of spatial resolution caused by pooling operation.

[0075] Among them, Figure 3 As shown in the figure, the dual attention block includes position attention and channel attention. The feature maps learned by the convolutional layer are used to extract spatial dependencies on position attention and channel dependencies on channel attention. The specific operation process is as follows:

[0076] The position attention generates three feature maps Q, K and V through three convolutions on the feature map, and then reshapes the three feature maps Q, K and V with dimensions C×W×H respectively to obtain three feature maps Q, K and V with dimensions C×N. Q is transposed and matrix multiplied with K and the softmax function is executed to obtain the position attention map A. The position attention map A is transposed and matrix multiplied with V, then reshaped to C×W×H and added and fused with the original feature map to obtain the position attention feature map Ep.

[0077] Channel attention directly reshapes the original feature map of dimension C×W×H three times to obtain three feature maps Q', K' and V' of dimension C×N, transposes Q', performs matrix multiplication with K' and executes softmax function to obtain channel attention map A', transposes channel attention map A', performs matrix multiplication with V', reshapes it to C×W×H and then adds and fuses it with the original feature map to obtain channel attention feature map Ec.

[0078] The position attention feature map Ep and the channel attention feature map Ec are fused using an adaptive fusion method. The adaptive fusion method of the two attention feature maps can fully capture the contextual information and enrich the representation of the convolutional features.

[0079]

[0080] Table 2

[0081] Among them, a transposed convolution with two times upsampling is first used, and then a transposed convolution with four times upsampling is used to enlarge the size of the second feature map to the target RGB image size. At the same time, a convolution layer with a stride of 1 and a size of 5×5 is used to reduce the checkerboard artifacts caused by the transposed convolution.

[0082] Step S600: Fusing the first feature map and the second feature map to obtain a target feature map, and locating an image modification region of the target RGB image based on the target feature map.

[0083] Specifically, the first feature map and the second feature map are first spliced ​​by depth to obtain a fused feature map, and then the fused feature map is used to obtain three feature maps of different sizes using convolution filters of 1×1, 3×3, and 5×5 respectively. The three feature maps of different sizes are then spliced ​​by depth to obtain a target feature map. Finally, logistic regression is used to perform pixel-by-pixel prediction on the target feature map to locate the image modification area of ​​the target RGB image.

[0084] Reference Figure 4 The embodiment of the present invention further provides an image modification region positioning system based on deep learning, wherein the image modification region positioning system based on deep learning includes:

[0085] The image acquisition unit 100 is used to acquire a target RGB image;

[0086] A first pre-filtering unit 200, used to extract an image residual of a target RGB image;

[0087] A first feature extraction unit 300 is used to extract inconsistency features of high-frequency components between a modified area and an unmodified area in a target RGB image based on an image residual using a first neural network, and generate a first feature map according to the extracted inconsistency features;

[0088] A second pre-filtering unit 400 is used to extract local binary pattern features of the target RGB image;

[0089] A second feature extraction unit 500 is used to merge the local binary pattern feature with the target RGB image to obtain a merged image, extract block similarity features between the modified area and the unmodified area in the merged image using a second neural network, and generate a second feature map according to the extracted block similarity features;

[0090] The feature fusion unit 600 is used to fuse the first feature map and the second feature map to obtain a target feature map, and locate the image modification area of ​​the target RGB image based on the target feature map.

[0091] It should be noted that since the deep learning-based image modification area positioning system in this embodiment and the above-mentioned deep learning-based image modification area positioning method are based on the same inventive concept, the corresponding contents in the method embodiment are also applicable to the present system embodiment and will not be described in detail here.

[0092] The embodiment of the present invention further provides a computer-readable storage medium, which stores computer-executable instructions. The computer-executable instructions are executed by one or more control processors, so that the one or more control processors can execute the image modification area positioning method based on deep learning in the above method embodiment, for example, executing the above described Figure 1 The functions of method steps S100 to S600 in the embodiment.

[0093] The system embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, i.e., may be located in one place, or may be distributed over multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0094] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform. Those skilled in the art can understand that all or part of the processes in the above embodiment method can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and the program can include the process of the embodiment of the above method when executed. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0095] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0096] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above-mentioned implementation mode. Technical personnel familiar with the field can also make various equivalent deformations or substitutions under the shared conditions without violating the spirit of the present invention. These equivalent deformations or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. A method for locating image modification areas based on deep learning, characterized in that: The following steps are involved: Acquire a target RGB image, where the target RGB image is a modified RGB image; Extracting an image residual of the target RGB image; Based on the image residual, a first neural network is used to extract inconsistency features of high-frequency components between the modified area and the unmodified area in the target RGB image, and a first feature map is generated according to the extracted inconsistency features; wherein, Extracting inconsistency features of high-frequency components between modified areas and unmodified areas in the target RGB image through four ResNet blocks, wherein each of the ResNet blocks includes two bottleneck units, and each of the bottleneck units includes three consecutive convolutional layers and an identity shortcut connection; Extracting local binary pattern features of the target RGB image; The local binary pattern feature is merged with the target RGB image to obtain a merged image, a block similarity feature between the modified area and the unmodified area in the merged image is extracted using a second neural network, and a second feature map is generated according to the extracted block similarity feature; wherein, Performing a convolution operation on the merged image through two consecutive convolution blocks and three dual attention convolution blocks to obtain three feature maps; Early fusion of the three feature maps is performed to extract block similarity features between the modified region and the unmodified region in the merged image; The first feature map and the second feature map are fused to obtain a target feature map, and an image modification region of the target RGB image is located based on the target feature map.

2. The image modification area positioning method based on deep learning according to claim 1, characterized in that: Extracting the local binary pattern features of the target RGB image includes the steps of: using the local binary pattern LBP to extract the local binary pattern features of the target RGB image, wherein the calculation formula of the local binary pattern LBP is: Among them, (xc, yc) represents the center pixel, ic represents the grayscale value of the center pixel, and i p Represents the grayscale value of the neighboring pixels of the central pixel.

3. The image modification area positioning method based on deep learning according to claim 1, characterized in that: The step of generating a second feature map according to the extracted block similarity features comprises the steps of: A transposed convolution with two times upsampling and a transposed convolution with four times upsampling are used to enlarge the size of the second feature map to the size of the target RGB image; and a convolution layer with a stride of 1 and a size of 5×5 is used to reduce the checkerboard artifacts caused by the transposed convolution.

4. The image modification area positioning method based on deep learning according to claim 1, characterized in that: The step of extracting the image residual of the target RGB image comprises the following steps: Using three first-order high-pass filters with initialized kernels to perform convolution operations on the three channels of the target RGB image respectively, to obtain three convolution results; The three convolution results are concatenated according to depth to obtain a residual image.

5. The image modification area positioning method based on deep learning according to claim 1, characterized in that: The step of generating a first feature map according to the extracted inconsistency features comprises the following steps: The size of the first feature map is enlarged to the target RGB image size through two quadruple upsampling transposed convolutions, and a convolution layer with a stride of 1 and a size of 5×5 is used to reduce the checkerboard artifacts caused by the transposed convolution.

6. The image modification area positioning method based on deep learning according to claim 1, characterized in that: The step of fusing the first feature map and the second feature map to obtain a target feature map, and locating the image modification area of ​​the target RGB image based on the target feature map comprises the following steps: The first feature map and the second feature map are concatenated according to the depth to obtain a fused feature map; The fused feature map is respectively subjected to convolutional filters of sizes 1×1, 3×3, and 5×5 to obtain three feature maps of different sizes; The three feature maps of different sizes are spliced ​​according to depth to obtain the target feature map; Logistic regression is used to perform pixel-by-pixel prediction on the target feature map to locate the image modification area of ​​the target RGB image.

7. An image modification area positioning system based on deep learning, characterized in that: include: An image acquisition unit, used for acquiring a target RGB image, wherein the target RGB image is a modified RGB image; A first pre-filtering unit, used to extract an image residual of the target RGB image; A first feature extraction unit is used to extract inconsistency features of high-frequency components between the modified area and the unmodified area in the target RGB image based on the image residual using a first neural network, and generate a first feature map according to the extracted inconsistency features; wherein, Extracting inconsistency features of high-frequency components between modified areas and unmodified areas in the target RGB image through four ResNet blocks, wherein each of the ResNet blocks includes two bottleneck units, and each of the bottleneck units includes three consecutive convolutional layers and an identity shortcut connection; A second pre-filtering unit, used for extracting local binary pattern features of the target RGB image; The second feature extraction unit is used to merge the local binary pattern feature with the target RGB image to obtain a merged image, use a second neural network to extract block similarity features between the modified area and the unmodified area in the merged image, and generate a second feature map according to the extracted block similarity features; wherein, Performing a convolution operation on the merged image through two consecutive convolution blocks and three dual attention convolution blocks to obtain three feature maps; Early fusion of the three feature maps is performed to extract block similarity features between the modified region and the unmodified region in the merged image; The feature fusion unit is used to fuse the first feature map and the second feature map to obtain a target feature map, and locate the image modification area of ​​the target RGB image based on the target feature map.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the image modification area positioning method based on deep learning as described in any one of claims 1 to 6.

Images