Power equipment nameplate image tampering detection method based on double-channel feature fusion

Abstract

The application relates to a power equipment nameplate image tampering detection method based on double-channel feature fusion, which comprises the following steps: acquiring a power equipment nameplate RGB image; inputting the image into a visual perception branch and a frequency perception branch of a power equipment nameplate image tampering detection model; the visual perception branch outputs visual features, and the frequency perception branch outputs frequency domain features; a hollow space pyramid pooling module of the power equipment nameplate image tampering detection model performs multi-scale fusion on the visual features and the frequency domain features to obtain fusion features; the fusion features are input into a decoder of the power equipment nameplate image tampering detection model, and a positioning result of a tampering region is output. Compared with the prior art, the application has the advantages of enhancing the accuracy and robustness of tampering region positioning.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion. Background Technology

[0002] With the intelligent development of power systems, the security of power equipment nameplate images, as crucial data in power production, is paramount. However, with advancements in image editing technology, the text and icons in power equipment nameplate images are easily maliciously tampered with, posing a serious threat to power information security. Existing image tampering detection technologies primarily target natural scene images and struggle to effectively detect minute text tampering areas in power equipment nameplate images. This is mainly because the nameplate material is prone to reflections, dirt, and other noise, and the tampered areas are typically small, making it difficult for existing general tampering detection algorithms to accurately identify them.

[0003] Existing image tampering detection methods mainly rely on the edge or surface features of natural images. However, the tampered areas in power equipment nameplate images usually have high visual consistency with the original image, making them difficult to detect using traditional visual features. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion. By combining visual and frequency domain features, the method improves the sensitivity to complex tampering traces (such as local erasure and forged textures). A ring residual module is designed to solve the problem of detail loss caused by feature degradation in deep networks. By combining multi-scale fusion and large kernel attention mechanism, the accuracy and robustness of tampering area localization are enhanced.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion, the method comprising the following steps:

[0007] The RGB image of the nameplate of the power equipment is acquired and input into the visual perception branch and frequency perception branch of the power equipment nameplate image tampering detection model. The visual perception branch outputs visual features and the frequency perception branch outputs frequency domain features.

[0008] The hollow spatial pyramid pooling module of the power equipment nameplate image tampering detection model performs multi-scale fusion of the visual features and frequency domain features to obtain fused features.

[0009] The decoder of the power equipment nameplate image tampering detection model, which integrates feature inputs, outputs the location result of the tampered area.

[0010] Furthermore, the image is input into the convolutional layer of the visual perception branch of the power equipment nameplate image tampering detection model. The convolutional layer outputs low-level features. The low-level features and high-level semantic features are input into the circular connection structure of the visual perception branch. The circular connection structure outputs visual features.

[0011] Furthermore, the specific steps for the circular connection structure to output visual features, where low-level features and high-level semantic features are input into the circular connection structure, are as follows:

[0012] Low-level features are processed through two convolutional and activation modules in a circular connection structure, and then through the max pooling layer of the circular connection structure in the visual perception branch. The features output by the max pooling layer are added to the low-level features to generate high-level semantic features. The low-level features and high-level semantic features are input into the circular connection structure, and the circular connection structure outputs visual features.

[0013] Furthermore, the Sobel high-pass filter module of the frequency sensing branch of the power equipment nameplate image tampering detection model is input into the image. The high-pass filter module outputs a residual image, which is then input into the ring connection structure of the frequency sensing branch. The frequency sensing branch outputs frequency domain features. The ring connection structure of the frequency sensing branch is the same as the ring connection structure of the visual sensing branch.

[0014] Furthermore, the void space pyramid pooling module includes four parallel void convolutional layers, each with a different void ratio.

[0015] Furthermore, the fusion characteristics are:

[0016]

[0017] Among them, F in For visual features and frequency domain features, Conv represents convolution, and DilatedConv represents... i This represents the i-th dilated convolutional layer.

[0018] Furthermore, the specific steps for outputting the location result of the tampered region in the decoder of the fusion feature input power equipment nameplate image tampering detection model are as follows:

[0019] The decoder of the power equipment nameplate image tampering detection model is upsampled layer by layer, and then concatenated with low-level features through skip connections. The concatenated features are input into the large kernel attention mechanism of the decoder, and the output is a weighted feature map. The weighted feature map is binarized to generate a tampered region mask, and the tampered region is located based on the mask.

[0020] Furthermore, the weighted feature map is the product of the concatenated features and the spatial-channel attention weight matrix of the large kernel attention mechanism.

[0021] Furthermore, the specific steps for obtaining the location result of the tampered area based on the mask are as follows:

[0022] The mask is denoised, and then candidate regions with areas greater than a threshold are selected by analyzing the connected components of the denoised mask as the location results of the tampered region.

[0023] Furthermore, the denoising employs morphological opening operations.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] This invention achieves bidirectional interaction between shallow details and deep semantics through the fusion of visual and frequency domain features, combined with a ring residual module, effectively solving the problems of high-frequency information loss and feature degradation in traditional methods. It utilizes a hollow spatial pyramid pooling (ASPP) module to capture multi-scale contextual information and enhances the decoder's ability to locate tampered region boundaries through a large kernel attention mechanism, significantly improving detection accuracy and anti-interference capabilities. Furthermore, this method supports multi-resolution input and possesses efficient inference performance, making it scalable for tamper detection of power equipment nameplates and other document-type images in complex scenarios. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the model structure of the present invention;

[0027] Figure 2 This is a flowchart of the present invention;

[0028] Figure 3 This is a structural diagram of the circular residual module (including residual propagation and feedback paths);

[0029] Figure 4 This is a schematic diagram of multi-scale dilated convolution in the ASPP module.

[0030] Figure 5 This is a diagram illustrating the working principle of the large-core attention mechanism.

[0031] Figure 6 This is a comparison chart of tamper detection results in the example, where... Figure 6 (a) is a tampered image. Figure 6 (b) is the test result. Figure 6 (c) represents the tampered area. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0033] This invention proposes a method for detecting tampering in power equipment nameplate images based on dual-channel feature fusion. By combining visual and frequency domain features, it can effectively detect minute tampering regions in power equipment nameplate images. Specifically, the method includes the following steps: First, visual features of the RGB image are extracted through a visual perception branch to capture color and texture information; second, frequency domain features of the image are extracted through a frequency perception branch to generate a residual image, capturing edges and subtle changes; next, the dual-channel features are fused at multiple scales using a hollow spatial pyramid pooling (ASPP) module to enhance the model's ability to perceive tampering regions at different scales; finally, the fused features are input into a decoder with a large kernel attention mechanism to decode layer by layer and accurately locate the tampering region. This invention has been experimentally validated on multiple international standard datasets. Experimental results show that the proposed method outperforms state-of-the-art algorithms in terms of precision, recall, and F-measure, exhibiting high accuracy and robustness, and is suitable for tampering detection and localization of equipment nameplate images in power systems.

[0034] like Figure 2 As shown, this invention proposes a method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion, comprising the following steps:

[0035] S1. Visual features of RGB images are extracted through a visual perception branch. The visual perception branch includes a circular residual module, which is used to transfer low-level features of the shallow network to the deep network and feed back high-level semantic features of the deep network to the shallow network.

[0036] S2. Extract the frequency domain features of the image through the frequency sensing branch. The frequency sensing branch generates a residual image through a high-pass filter and uses a ring residual module to extract the edge and subtle change features in the residual image.

[0037] S3. Multi-scale fusion of the visual features and frequency domain features is performed using the Spatial Pyramid Pooling (ASPP) module, wherein the ASPP module contains multiple parallel dilated convolutional layers with different dilation rates.

[0038] S4. Input the fused features into the decoder, which is equipped with a large kernel attention mechanism. Through layer-by-layer decoding and contextual information enhancement, the decoder outputs the localization result of the tampered region.

[0039] The circular residual module implements feature interaction in the following ways:

[0040] S11. The low-level features output by the shallow network are passed to the deep network through the residual propagation path;

[0041] S12. Feed the high-level semantic features output by the deep network to the shallow network through the residual feedback path;

[0042] S13. The residual propagation path and the residual feedback path are integrated through a ring connection structure to form a closed-loop feature interaction.

[0043] The high-pass filter is a Sobel operator, and its convolution kernel is:

[0044]

[0045] Where S x and S y These are represented as filters in the vertical and horizontal directions, respectively.

[0046] The porosity of the ASPP module includes 1, 6, 12, and 18, and the kernel size of each dilated convolutional layer...

[0047] The large kernel attention mechanism uses a convolutional kernel size of 9×9 to 15×15 to calculate the attention weight matrix and perform weighted fusion of the feature maps of the decoder based on the weight matrix.

[0048] The decoder outputs the location result of the tampered area through the following steps:

[0049] S41. The fused feature map is sequentially passed through an upsampling layer and a skip connection to be concatenated with the low-level features.

[0050] S42. Capture cross-scale contextual information through a large kernel attention mechanism;

[0051] S43. Binarize the weighted feature map to generate a tampered region mask.

[0052] The method also includes post-processing of the tampered region mask, including:

[0053] S51. Remove noise interference through morphological operations;

[0054] S52. Use connected component analysis to filter regions with areas greater than a preset threshold as the final tampered regions.

[0055] Compared with existing technologies, this invention achieves bidirectional interaction between shallow details and deep semantics by fusing visual and frequency domain features and combining them with a ring residual module, effectively solving the problems of high-frequency information loss and feature degradation in traditional methods. It captures multi-scale contextual information using a hollow spatial pyramid pooling (ASPP) module and enhances the decoder's ability to locate tampered region boundaries through a large kernel attention mechanism, significantly improving detection accuracy and anti-interference capabilities. Furthermore, this method supports multi-resolution input and possesses efficient inference performance, making it scalable for tamper detection of power equipment nameplates and other document-type images in complex scenarios.

[0056] like Figure 1 As shown, the tampering detection method of the present invention includes an image preprocessing module, a dual-channel feature extraction module, a multi-scale feature fusion module, a decoder and attention module, and a post-processing module.

[0057] ① Image preprocessing module: responsible for distortion correction, size normalization and pixel value standardization of the input image;

[0058] ② Dual-channel feature extraction module: includes a visual perception branch and a frequency perception branch, which extract the visual features and frequency domain features of the image respectively;

[0059] ③ Multi-scale fusion module: Multi-scale feature fusion is achieved through hollow spatial pyramid pooling (ASPP);

[0060] ④ Decoder and Attention Module: Based on the big-kernel attention mechanism, accurately locate the tampered area;

[0061] ⑤ Post-processing module: performs noise filtering and region selection on the detection results.

[0062] This invention includes the following steps:

[0063] First, after acquiring the RGB image of the power equipment nameplate, affine transformation is used to correct geometric distortion, with rotation angle controlled within ±5° and scaling ratio limited to 0.9 to 1.1 times. Bilinear interpolation is then used to uniformly scale the input image to 512×512 pixels, and Min-Max normalization is performed on the RGB channels, mapping pixel values ​​to the [0,1] interval. For input images with variable resolution, dynamic resizing ensures consistent processing.

[0064] Then, the image is input into the dual-channel feature extraction module. The visual perception branch uses a circular residual module, such as... Figure 3 As shown. This module transmits shallow, low-level features (such as edges and textures) to the deep network through a residual propagation path, while simultaneously transmitting deep, high-level semantic features back to the shallow layer through a residual feedback path, forming a closed-loop feature interaction. This addresses the feature degradation problem in deep networks and outputs a 32×32×256 feature map. The frequency-aware branch uses a Sobel filter (with vertical convolution kernels as follows:

[0065]

[0066] The horizontal direction is:

[0067]

[0068] A residual image is generated to capture high-frequency tampering traces. After the residual image is optimized by 3 layers of convolution (kernel size 3×3) and a ring residual module, a 32×32×256 frequency domain feature map is output.

[0069] Input the dual-channel features into the Spatial Pyramid Pooling (ASPP) module (e.g.) Figure 4 (As shown). This module contains four parallel dilated convolutional layers (with dilation rates of 1, 6, 12, and 18, and kernel size of 3×3), which fuse multi-scale contextual information through weighted summation to output a 32×32×512 fused feature map. The formula is expressed as:

[0070]

[0071] The fused features are upsampled layer by layer by the decoder (stride 2, deconvolution kernel size 4×4) and concatenated with shallow features through skip connections. A large kernel attention mechanism (convolution kernel size 11×11) is embedded in the decoder to generate a spatial-channel attention weight matrix, which weights and enhances the response of the tampered region. Finally, a 512×512 tamper mask is generated through Sigmoid activation and binarization (threshold 0.5).

[0072] Finally, morphological opening operations (kernel size 3×3) are performed on the mask to remove isolated noise, and candidate regions with an area ≥ 50 pixels are selected as the final tampering regions through connected component analysis (e.g., Figure 5 (As shown).

[0073] Comparison of tampered test results (as shown in the image) Figure 6 As shown.

[0074] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion, characterized in that, The method includes the following steps: The RGB image of the nameplate of the power equipment is acquired and input into the visual perception branch and frequency perception branch of the power equipment nameplate image tampering detection model. The visual perception branch outputs visual features and the frequency perception branch outputs frequency domain features. The hollow spatial pyramid pooling module of the power equipment nameplate image tampering detection model performs multi-scale fusion of the visual features and frequency domain features to obtain fused features. The decoder of the power equipment nameplate image tampering detection model, which integrates feature inputs, outputs the location result of the tampered area.

2. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 1, characterized in that, The image is input into the convolutional layer of the visual perception branch of the power equipment nameplate image tampering detection model. The convolutional layer outputs low-level features. The low-level features and high-level semantic features are input into the circular connection structure of the visual perception branch. The circular connection structure outputs visual features.

3. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 2, characterized in that, The specific steps by which low-level features and high-level semantic features are input into the circular connection structure, and the circular connection structure outputs visual features are as follows: Low-level features are processed through two convolutional and activation modules in a circular connection structure, and then through the max pooling layer of the circular connection structure in the visual perception branch. The features output by the max pooling layer are added to the low-level features to generate high-level semantic features. The low-level features and high-level semantic features are input into the circular connection structure, and the circular connection structure outputs visual features.

4. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 3, characterized in that, The Sobel high-pass filter module of the frequency sensing branch of the power equipment nameplate image tampering detection model is input into the image. The high-pass filter module outputs a residual image. The residual image is input into the ring connection structure of the frequency sensing branch. The frequency sensing branch outputs frequency domain features. The ring connection structure of the frequency sensing branch is the same as the ring connection structure of the visual sensing branch.

5. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 1, characterized in that, The void space pyramid pooling module consists of four parallel voided convolutional layers, each with a different void ratio.

6. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 1, characterized in that, The fusion features are: Among them, F in For visual features and frequency domain features, Conv represents convolution, and DilatedConv represents... i This represents the i-th dilated convolutional layer.

7. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 6, characterized in that, The specific steps for outputting the location result of the tampered region in the decoder of the tampering detection model for power equipment nameplate images with fused features are as follows: The decoder of the power equipment nameplate image tampering detection model is upsampled layer by layer, and then concatenated with low-level features through skip connections. The concatenated features are input into the large kernel attention mechanism of the decoder, and the output is a weighted feature map. The weighted feature map is binarized to generate a tampered region mask, and the tampered region is located based on the mask.

8. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 7, characterized in that, The weighted feature map is the product of the concatenated features and the spatial-channel attention weight matrix of the big kernel attention mechanism.

9. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 8, characterized in that, The specific steps to obtain the location result of the tampered area based on the mask are as follows: The mask is denoised, and then candidate regions with areas greater than a threshold are selected by analyzing the connected components of the denoised mask as the location results of the tampered region.

10. The method for detecting tampering of power equipment nameplate images based on dual-channel feature fusion according to claim 9, characterized in that, The denoising process employs morphological opening operations.

Images