Depth decision fusion tampered image positioning method based on optimization U-Net
Through the deep decision-making fusion method based on optimized U-Net, the U-Net network is constructed, which solves the problem of poor image tampering positioning effect in the existing technology, and accurately locates and recognizes the image tampering area, and improves generalization capabilities and visual effects.
Patent Information
- Application Number
- CN202411494239.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2024-12-27
AI Technical Summary
When judging the authenticity of digital images, the positioning effect is not ideal, the algorithm generalization ability is poor, and it is difficult to accurately identify and locate the image tampering area.
U-Net network is built using a deep decision fusion method based on optimization of U-Net, and the network is trained through data sets generated by a variety of tampering tools. The U-Net architecture is built using encoder, decoder and output layer. It includes a convolutional layer, a ReLU activation function, a maximum pooling layer and a Dropout layer to realize the tampering area positioning of the image.
It realizes accurate positioning and identification of image tampering areas, improves visual effects and generalization capabilities, and is suitable for applications in multiple scenarios. It can effectively distinguish the tampered parts from the original images, and maintains the clarity and consistency of the positioning map.
Smart Images
Figure CN119205894A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a method for locating tampered images based on a deep decision fusion optimized U-Net. Background Art
[0002] Digital images are widely used in daily life and various fields, such as social media, news dissemination, medical imaging, and security monitoring. However, with the development of image processing software, it is very easy to tamper with digital images.
[0003] The existing methods for judging the authenticity of digital images usually only use a single trace, and it is necessary to consider the output results of multiple complementary algorithms. In recent years, scholars at home and abroad have proposed research on tampered images using decision fusion algorithms, but the positioning effect is not ideal and the algorithm generalization ability is poor. Summary of the invention
[0004] The technical problem of the present invention is: to seek a tampered image positioning method based on a deep decision fusion optimized U-Net, to build a network with a U-Net architecture in combination with deep decision fusion, the network integrates data from multiple tampered image positioning tools, and compares the located tampered image with the real image and mask to further judge the authenticity of the image and the positioning effect, while having a high generalization ability.
[0005] The purpose of the present invention is to solve the above problems and propose a method for tampered image localization based on deep decision fusion optimized U-Net, which includes the following steps: S1: The collected images are tampered with using a variety of image tools to construct a synthetic dataset D1, and the synthetic dataset D1 is divided into a training set and a test set; S2: Construct U-Net network by combining deep decision fusion; S3: Input the training set of the synthetic dataset D1 into the U-Net network for training; S4: Input the test set of the synthetic dataset D1 into the U-Net network and locate the tampered area in the image based on the output results. Furthermore, in step S1, the collected images are tampered with using a variety of image tools, including tampering with the collected images by splicing, copying, moving and removing, specifically rotating, resizing and replacing a certain area of the image.
[0006] Preferably, the collected images are collected through different acquisition devices, and the storage format of the images is not restricted. The collected images are collected through different tampering monitoring tools to construct a synthetic data set.
[0007] Preferably, the U-Net network combined with deep decision fusion includes an encoder, a decoder and an output layer, the encoder includes two types of encoding blocks, and the decoder includes multiple decoding blocks.
[0008] Furthermore, the encoder includes two types of encoding blocks, the first type includes multiple convolutional layers, ReLU activation function layers and maximum pooling layers, and the second type of encoding blocks includes multiple convolutional layers, ReLU activation function layers, maximum pooling layers and Dropout layers.
[0009] Preferably, the Dropout layer is used to reset the output of some neurons to zero, expressed as: ; In the formula, represents the convolutional layer output of the encoding block, Represents the output result of the Dropout layer.
[0010] Furthermore, the decoding block includes an upsampling layer, a convolutional layer, a ReLU activation function layer, and a connection layer.
[0011] Preferably, the connection layer connects the convolution output and the full-resolution output of the encoding part to reconstruct the full-resolution feature fusion map, which is expressed as:
[0012] In the formula, represents the convolution output of the decoding part, Indicates the full-resolution output of the encoding part, Represents the convolution output.
[0013] Preferably, the output layer includes a convolution, ReLU activation function layer and a convolution, Sigmoid activation function layer.
[0014] Compared with the prior art, the beneficial effects of the present invention include: 1) This method proposes a deep decision fusion tampered image localization method based on optimized U-Net, fuses the output maps generated by multiple tampering tools to construct a synthetic dataset, and trains the U-Net network with the synthetic dataset. Due to the fusion of the output maps generated by multiple tampering localization tools, the tampered area can be accurately marked.
[0015] 2) The U-Net constructed by the present invention in combination with deep decision fusion has good visual effects, and the U-Net network structure is simple and has high generalization ability, which is suitable for use in various scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0017] Figure 1A flowchart of tampering image location according to an embodiment of the present invention; Figure 2 This is a flowchart of image tampering of a synthetic data set according to an embodiment of the present invention; Figure 3 A network structure diagram of optimizing U-Net according to an embodiment of the present invention; Figure 4 The figure is a schematic diagram of a tampered picture, a real picture and a mask according to an embodiment of the present invention. DETAILED DESCRIPTION
[0018] like Figure 1 As shown, the deep decision fusion tampered image positioning method based on optimized U-Net includes the following steps: S1: The collected images are tampered with using a variety of image tools to construct a synthetic dataset D1, and the synthetic dataset D1 is divided into a training set, a test set, and a validation set; As shown in Table 1, the collected images are collected through different acquisition devices, and the storage format of the images is not restricted. The collected images are collected through different tampering monitoring tools to construct a synthetic dataset. The tampering monitoring tools include FaceSwaps, IEEE IFS, IMD2020 and Korus. The synthetic dataset contains 3200 tampered images and their corresponding tampering masks.
[0019] Table 1
[0020] like Figure 2 As shown, the collected images are tampered with, including splicing, copying, moving and removing, specifically rotating, resizing and replacing a certain area of the image; the steps of image tampering are to randomly replace an area of the original image with a JPEG compressed image, and the compressed image is replaced by rotating, resizing and adjusting the quality. The rotation interval is in the range of [-10, 10], the resizing ratio coefficient is in the range of [0.75, 1.25], and the quality coefficient is any coefficient in (-70, 75, 80, 85, 90, 95, 100).
[0021] S2: Construct a deep decision fusion tampered image localization method based on optimized U-Net; The U-Net network is constructed in combination with deep decision fusion, including an encoder, a decoder and an output layer. The encoder contains two types of encoding blocks, and the decoder contains multiple decoding blocks.
[0022] The encoder contains two types of encoding blocks. The first type contains multiple convolutional layers, ReLU activation function layers and maximum pooling layers, and the second type of encoding blocks includes multiple convolutional layers, ReLU activation function layers, maximum pooling layers and Dropout layers.
[0023] like Figure 3 As shown, the encoder includes 2 first-category coding blocks and 1 second-category coding block.
[0024] For the input of the training set of the dataset, the first type of encoding block is first executed twice Convolution, then execute Maximum pooling is used to compress the information contained in the input; the dropout layer is used to reset the output of some neurons to zero to improve the generalization ability of the network; the number of filters is 16, 32, and 64 respectively.
[0025] The calculation formula of the Dropout layer is: ; In the formula, represents the convolutional layer output of the encoding block, Represents the output result of the Dropout layer.
[0026] The decoding part contains 3 decoding blocks. Each decoding block upsamples the input map and performs 3×3 convolution to double the dimension. The output of the convolution is connected to the full-resolution output of the corresponding encoding block through the connection layer. The combined map is then passed through two layers. The number of filters in the decoding block is 64, 32, and 16 respectively.
[0027] The calculation formula of the connection layer is:
[0028] In the formula, represents the convolution output of the decoding part, Indicates the full-resolution output of the encoding part, Represents the convolution output.
[0029] The output layer includes convolution, ReLU activation function layer and convolution, Sigmoid activation function layer.
[0030] S3: Input the training set of data set D1 into the network for training.
[0031] S4: Input the test set of dataset D1 into the network and locate the tampered area in the image based on the output results. like Figure 4As shown, three types of tampered images, namely, splicing, copying and moving, and removing, and real pictures of the three types of tampered images and corresponding tampered positioning maps of the three types of tampered images generated by this method, the tampered positioning map is a mask; the three groups of pictures prove that the application of the embodiments of the present invention has the following effects.
[0032] Clear edges: The edges of each tampered area are very clear, effectively distinguishing the tampered part from other areas of the original image and enhancing the readability of the positioning map.
[0033] Precise positioning: Regardless of splicing, copying, moving or removing, the present invention accurately identifies and locates the tampered area, achieving the accuracy and stability of locating the tampered area of the network of the present invention.
[0034] Consistency: In different types of tampered images, the generation of masks remains consistent, so that the network of the present invention has good adaptability to various tampering methods.
[0035] Feature extraction: The network of the present invention effectively extracts the tampering information in the image, so that the tampering location map can clearly reflect the nature of the tampered content.
Claims
1. A deep decision fusion tampered image localization method based on optimized U-Net, characterized by: The method comprises: S1: The collected images are tampered with using a variety of image tools to construct a synthetic dataset D1, and the synthetic dataset D1 is divided into a training set and a test set; S2: Construct U-Net network by combining deep decision fusion; S3: Input the training set of the synthetic dataset D1 into the network for training; S4: Input the test set of the synthetic dataset D1 into the network, and locate the tampered area in the image based on the network output results.
2. According to claim 1, the method for tampered image positioning based on deep decision fusion optimized U-Net is characterized in that: In step S1, the collected images are tampered with using a variety of image tools, including tampering with the collected images by splicing, copying, moving and removing, specifically rotating, resizing and replacing a certain area of the image.
3. According to claim 1, the method for tampered image positioning based on deep decision fusion optimized U-Net is characterized in that: In step S2, the U-Net network is constructed by combining deep decision fusion, including an encoder, a decoder and an output layer, the encoder includes two types of encoding blocks, and the decoder includes multiple decoding blocks.
4. According to claim 3, the method for tampered image localization based on deep decision fusion optimized U-Net is characterized in that: The two types of coding blocks, the first type includes multiple convolutional layers, ReLU activation function layers and maximum pooling layers, and the second type of coding blocks includes multiple convolutional layers, ReLU activation function layers, maximum pooling layers and Dropout layers.
5. According to claim 4, the method for tampered image positioning based on deep decision fusion optimized U-Net is characterized in that: The Dropout layer is used to return the output of some neurons to zero, and the expression is: ； In the formula, represents the convolutional layer output of the encoding block, Represents the output result of the Dropout layer.
6. According to claim 3, the method for tampered image positioning based on deep decision fusion optimized U-Net is characterized in that: The decoding block includes an upsampling layer, a convolution layer, a ReLU activation function layer and a connection layer.
7. According to claim 6, the method for tampered image localization based on deep decision fusion optimized U-Net is characterized in that: The connection layer connects the convolution output and the full-resolution output of the encoding part to reconstruct the full-resolution feature fusion map, and the expression is: In the formula, represents the convolution output of the decoding part, Indicates the full-resolution output of the encoding part, Represents the convolution output.
8. According to claim 3, the method for tampered image positioning based on deep decision fusion optimized U-Net is characterized in that: The output layer includes a convolution, ReLU activation function layer and a convolution, Sigmoid activation function layer.