Neural network-based forgery detection method and system

Abstract

The invention provides a counterfeit detection method and system based on a neural network, and the method comprises the steps: carrying out the depth feature extraction of a frequency domain feature, a texture feature and a time domain feature through employing a preset enhanced FPN network, so as to obtain a depth spectrum feature, a depth texture feature and a depth time domain feature; and calculating a comprehensive score according to the acquired abnormal frequency domain, abnormal texture description, abnormal smooth transition frame and a preset module weight so as to judge whether the target sample is forged or not, and performing inverse feedback based on a judgment result so as to acquire the use frequency of each neuron in the enhanced FPN network. And if the use frequency of the neurons is lower than a preset use threshold, soft pruning is performed on the neurons to update the reinforced FPN network, so that the reinforced FPN network is reinforced during use, the reinforced FPN network is continuously optimized, the precision and the effectiveness of the reinforced FPN network are improved, the conditions of forgery detection are reduced, and the difficulty of forgery detection is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image and video processing, and more specifically to a forgery detection method and system based on a neural network. Background Art

[0002] In the prior art, when identifying image and video forgeries, deep learning methods are usually used to automatically learn and extract features using models such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM). The method based on generative adversarial networks (GAN) uses the game characteristics between the generator and discriminator of GANs to enhance the sensitivity of the detector to deep forgeries. The multimodal fusion method is used to combine multimodal information such as vision, audio, and semantics to improve the accuracy and robustness of detection. The method based on physical constraints analyzes physical properties such as lighting, shadows, and reflections to determine whether the content conforms to the physical laws of the real world. The method based on biological signals analyzes biological signals such as heartbeat and breathing rate because these signals are difficult to accurately imitate.

[0003] But there are the following defects:

[0004] (1) Complex scenes and lighting conditions may make it difficult to distinguish the physical characteristics of real data from forged data, requiring additional sensor data, which limits the general applicability of this method and places high demands on equipment;

[0005] (2) The synchronization and fusion technology of different modal data still needs to be improved, and the collection and processing costs of multimodal data are relatively high;

[0006] (3) GANs themselves may be used to create more advanced deep fakes, making detection more difficult;

[0007] (4) Some complex methods may not be able to meet the requirements of real-time detection, especially on resource-constrained devices.

[0008] Therefore, there is an urgent need for a neural network-based counterfeit detection method and system that can deeply capture detailed information, does not require the use of sensor data, reduces detection difficulty, reduces equipment reliance, and improves counterfeit identification accuracy. Summary of the invention

[0009] In view of the above problems, the purpose of the present invention is to provide a counterfeit detection method and system based on a neural network to solve the technical problems that the prior art requires additional sensor data, which limits the universal applicability of this method, and has high equipment requirements, high processing costs, high detection difficulty, and easily restricted detection.

[0010] The present invention provides a neural network-based forgery detection method, which includes:

[0011] The target sample is subjected to feature extraction by a preset frequency domain feature extraction module, a texture feature extraction module and a time domain feature extraction module to obtain frequency domain features, texture features and time domain features respectively;

[0012] Using a preset enhanced FPN network to perform deep feature extraction on the frequency domain features, the texture features and the time domain features to obtain deep spectrum features, deep texture features and deep time domain features;

[0013] Based on the deep spectrum feature, the abnormal frequency domain is locked, based on the deep texture feature, the abnormal texture description is captured, based on the deep time domain feature, the abnormal smooth transition frame is acquired, and a comprehensive score is calculated according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weight;

[0014] Comparing the comprehensive score with a preset forgery threshold to determine whether the target sample is a forged sample and forming a determination result;

[0015] Based on the determination result, reverse feedback is performed to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of a neuron is lower than a preset usage threshold, soft pruning is performed on the neuron to update the enhanced FPN network.

[0016] Optionally, the latest updated enhanced FPN network is adopted.

[0017] Optionally, the target sample is a target image or a target video; wherein,

[0018] If the target sample is a target image, the time domain feature and the depth time domain feature are empty.

[0019] Optionally, the locking of the abnormal frequency domain based on the deep spectrum feature includes:

[0020] Performing spectrum analysis according to the deep spectrum features to form spectrum analysis data;

[0021] Decomposing the image information of the target sample into gradient divisions of different frequencies based on the spectrum analysis data, and forming a visible spectrum based on the gradient divisions;

[0022] A prominent area above or below a preset spectrum threshold is locked on the visible spectrum, and a region corresponding to the prominent area on the target sample is taken as an abnormal region.

[0023] Optionally, capturing abnormal texture description based on the deep texture feature includes:

[0024] Acquire a unit pixel value and adjacent pixel values ​​of the unit pixel in the target sample based on the depth texture feature;

[0025] Calculate a local binary pattern feature value based on the unit pixel value and the adjacent pixel values;

[0026] The local binary pattern feature value is compared with a preset texture threshold to capture abnormal texture.

[0027] Optionally, acquiring an abnormally smooth transition frame based on the deep time domain feature includes:

[0028] Determine whether the depth time domain feature is empty; wherein,

[0029] If it is empty, output 0;

[0030] If it is not empty, inter-frame difference detection is performed according to the deep time domain features to obtain the pixel difference between adjacent frames, and the pixel difference is compared with a preset smoothing threshold. If the pixel difference is greater than the smoothing threshold, the two adjacent frames are regarded as abnormal smooth transition frames.

[0031] Optionally, calculating a comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and a preset module weight includes:

[0032] Calculate the frequency domain ratio of the abnormal frequency domain in the target sample, the area ratio of the abnormal texture description in the target sample, and the total frame ratio of the abnormal smooth transition frame in the target sample;

[0033] The frequency domain proportion is multiplied by a preset magnification value, and then multiplied by a preset frequency domain feature extraction module weight value to obtain a frequency domain score; the area proportion is multiplied by the magnification value, and then multiplied by a preset texture feature extraction module weight value to obtain a texture score; and the total frame proportion is multiplied by the magnification value, and then multiplied by a preset time domain feature extraction module weight value to obtain a time domain score;

[0034] The frequency domain score, the texture score, and the time domain score are added together to obtain a comprehensive score.

[0035] Optionally, the use frequency of each neuron in the FPN network is enhanced and displayed in the determination result, including:

[0036] The preset forward propagation network obtains the usage frequency and amplitude data of the neuron based on the preset neuron counter, and writes the usage frequency and amplitude data into the determination result; wherein,

[0037] The neuron counter is preset after each neuron in the enhanced FPN network. When the output of the neuron exceeds a preset output threshold, the neuron counter is incremented by one and the amplitude data of the neuron is recorded.

[0038] The present invention also provides a forgery detection system based on a neural network, wherein the forgery detection method based on a neural network as described above is implemented, and the system includes a frequency domain feature extraction module, a texture feature extraction module, a time domain feature extraction module, an enhanced FPN network, a determination module, and a soft pruning module, wherein:

[0039] The frequency domain feature extraction module, the texture feature extraction module and the time domain feature extraction module are used to respectively extract features from the target sample to obtain frequency domain features, texture features and time domain features;

[0040] The enhanced FPN network is used to perform deep feature extraction on the frequency domain features, the texture features and the time domain features to obtain deep spectrum features, deep texture features and deep time domain features;

[0041] The determination module is used to lock the abnormal frequency domain based on the deep spectrum feature, capture the abnormal texture description based on the deep texture feature, obtain the abnormal smooth transition frame based on the deep time domain feature, and calculate the comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weight; and compare the comprehensive score with the preset forgery threshold to determine whether the target sample is a forged sample to form a determination result;

[0042] The soft pruning module is used to perform reverse feedback based on the determination result to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of a neuron is lower than a preset usage threshold, the neuron is soft pruned to update the enhanced FPN network.

[0043] Optionally, the enhanced FPN network that performs deep feature extraction on the frequency domain features, the texture features and the time domain features in real time to obtain deep spectrum features, deep texture features and deep time domain features is the latest updated enhanced FPN network.

[0044] From the above technical solution, it can be seen that the present invention provides a method and system for counterfeit detection based on a neural network. First, a preset frequency domain feature extraction module, a texture feature extraction module and a time domain feature extraction module are used to extract features of a target sample to obtain frequency domain features, texture features and time domain features respectively. Then, a preset enhanced FPN network is used to perform deep feature extraction on the frequency domain features, texture features and time domain features to obtain deep spectrum features, deep texture features and deep time domain features. Then, the abnormal frequency domain is locked based on the deep spectrum features, the abnormal texture description is captured based on the deep texture features, and the abnormal smooth transition frame is obtained based on the deep time domain features. The comprehensive score is calculated according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weights, so as to determine whether the target sample is counterfeit, and reverse feedback is performed based on the judgment result to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of the neuron is lower than the preset usage threshold, the neuron is soft pruned to update the enhanced FPN network, so as to strengthen the enhanced FPN network during use, thereby continuously optimizing the enhanced FPN network, improving the accuracy and effectiveness of the enhanced FPN network, reducing the conditions for counterfeit detection, and reducing the difficulty of counterfeit detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] By referring to the following description in conjunction with the accompanying drawings, and with a more comprehensive understanding of the present invention, other objects and results of the present invention will become more clear and easy to understand. In the accompanying drawings:

[0046] Figure 1 is a flowchart of a neural network-based counterfeit detection method according to an embodiment of the present invention;

[0047] Figure 2 A logical schematic diagram of a counterfeit detection method based on a neural network according to an embodiment of the present invention;

[0048] Figure 3 Schematic diagram of a neural network-based counterfeit detection system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] Existing methods for detecting forgeries in images or videos require additional sensor data, which limits the general applicability of this method. They also have high equipment requirements, high processing costs, high detection difficulty, and are easily restricted.

[0050] In view of the above problems, the present invention provides a method and system for counterfeit detection based on a neural network. The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] In order to illustrate the neural network-based counterfeit detection method and system provided by the present invention, Figure 1-Figure 3 The embodiments of the present invention are indicated by way of example.

[0052] The following description of exemplary embodiments is merely illustrative in nature and is not intended to limit the present invention and its application or use. Technologies and devices known to ordinary technicians in the relevant field may not be discussed in detail, but in appropriate cases, the technologies and devices should be considered as part of the specification.

[0053] like Figure 1 , Figure 2 As shown in the figure, the present invention provides a counterfeit detection method based on a neural network, comprising:

[0054] S1: extracting features of the target sample through a preset frequency domain feature extraction module, texture feature extraction module and time domain feature extraction module to obtain frequency domain features, texture features and time domain features respectively;

[0055] S2: using a preset enhanced FPN network to perform deep feature extraction on the frequency domain features, the texture features and the time domain features to obtain deep spectrum features, deep texture features and deep time domain features;

[0056] S3: locking the abnormal frequency domain based on the deep spectrum feature, capturing the abnormal texture description based on the deep texture feature, acquiring the abnormal smooth transition frame based on the deep time domain feature, and calculating the comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weight;

[0057] S4: comparing the comprehensive score with a preset forgery threshold to determine whether the target sample is a forged sample and forming a determination result;

[0058] S5: Based on the determination result, reverse feedback is performed to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of a neuron is lower than a preset usage threshold, soft pruning is performed on the neuron to update the enhanced FPN network.

[0059] In this embodiment, the target sample is a target image or a target video; if the target sample is an image (target image), it is a single-frame image, so there is no time domain. Therefore, if the target sample is a target image, the time domain feature and the depth time domain feature are empty.

[0060] In this embodiment, step S1 performs feature extraction on the target sample through a preset frequency domain feature extraction module, texture feature extraction module and time domain feature extraction module to obtain frequency domain features, texture features and time domain features respectively.

[0061] In a specific embodiment, the frequency domain feature extraction module is used to extract frequency domain features. For forgery detection, the frequency domain feature extraction module performs frequency domain analysis on deep forgery detection through Fast Fourier Transform (FFT), that is, by converting the image from the spatial domain to the frequency domain, to identify abnormal features that may exist in the image. For a given image I (x, y), its two-dimensional fast Fourier transform F (u, v) can be expressed as:

[0062]

[0063] Where M and N are the width and height of the image, u and v are frequency variables, and j is the imaginary unit.

[0064] In the context of deepfakes, these fake anomalies may arise from technical limitations or human errors in the image or video synthesis process, such as problems in boundary processing, resolution matching, or compression artifacts. These anomalies are highlighted by performing the FFT amplitude channel of the Fourier transform, making it easier to lock the abnormal area. When searching for anomalies in frequency domain features, frequency domain conversion is first performed. The Fourier algorithm is used to convert the image from the spatial domain to the frequency domain. In the frequency domain, the image information is decomposed into components of different frequencies, including low frequencies (corresponding to the smooth parts of the image) and high frequencies (corresponding to the details and edges of the image). Spectral analysis is then performed, that is, the spectrum of the image or video frame is checked in the frequency domain to find abnormal frequency components. For example, deep fake images may show unnatural patterns at specific frequencies, such as repeated textures at the image boundary, discontinuous edges, or inconsistent noise distribution. Spectral visualization is then performed, that is, by visualizing the spectrum of the image, the anomalies in the frequency domain can be intuitively observed. For example, if the forged image has too high energy at a certain frequency, this may indicate traces of forgery. Finally, feature extraction is performed to extract specific features in the frequency domain, such as energy distribution, discontinuity of frequency distribution, etc. These features can be used to train machine learning or deep learning models to identify forged images, etc.

[0065] In addition, when pre-training the frequency domain feature extraction module, inverse transform verification is also required, that is, using the Inverse Fast Fourier Transform (IFFT) to convert the frequency domain information back to the spatial domain. The details of the video frame of the image or video can be further analyzed to verify whether there are unnatural transitions or repetitive patterns, thereby improving the accuracy of anomaly detection.

[0066] In the process of texture feature extraction by the texture feature extraction module, a local binary pattern (LocalBinary Pattern, referred to as LBP) is often used for calculation to achieve image enhancement, improve robustness and calculation simplicity, and improve the effectiveness of capturing local texture information. Therefore, in this embodiment, the specific texture and edge information in the extracted texture features will be used to distinguish between the real image and the generated image, that is, each pixel can be compared with its adjacent pixels to obtain the texture features, thereby facilitating the locking of the abnormal area.

[0067] For the central pixel c and its neighboring pixels p i , LBP eigenvalue LBP C It can be defined as:

[0068]

[0069] in,

[0070]

[0071] This makes it easier to determine whether there is an abnormality in the area by comparing the difference between specific pixels and surrounding pixels.

[0072] In the process of extracting temporal features through the temporal feature extraction module, forgery detection mainly focuses on the continuity and consistency between frames in the video sequence. This part is only for the case where the target sample is a video. Since a simple image only has a single frame, there is no temporal feature. For video forgery prevention, the temporal detection technology uses the characteristics of video content changing over time to check the consistency of motion, appearance and structure between video frames.

[0073] Therefore, we can subsequently identify the inter-frame differences in the video by comparing the pixel differences between adjacent frames to capture abnormal editing points. For example, the pixel values ​​at the splicing points of a deep fake video may change suddenly, which is different from the smooth transition of a normal video. The formula for calculating the inter-frame difference is as follows:

[0074] D(x,y)=|F t (x,y)-F t+1 (x,y)|;

[0075] Among them, F t and F t+1 represents adjacent frames, (x, y) represents pixel coordinates, and |·| represents absolute values.

[0076] Step S2 is a process of using a preset enhanced FPN network to perform deep feature extraction on frequency domain features, texture features and time domain features to obtain deep spectrum features, deep texture features and deep time domain features. The enhanced FPN (Feature Pyramid Network) network is an enhanced version of the improved pyramid network, which extracts the frequency domain features, texture features and time domain features obtained above from bottom to top in a deeper way, such as through different levels C in the improved pyramid network. 1 ,C 2 ,…,C n The bottom-up pathway obtains features (feature maps) at different levels until the top-level feature map is obtained to form the corresponding deep spectrum features, deep texture features, and deep time domain features; where P n is the first (topmost) feature map in the top-down path, usually equal to C n , for P n-1 ,P n-2 ,…,P 1 The calculation formula is as follows:

[0077] P n =C n

[0078] For n>1:

[0079] P i =Conv 1×1 (C i +UpSample(P i+1 ))

[0080] Among them, Conv 1×1 It represents a 1×1 convolution operation, which is used to adjust the number of channels of the feature map. UpSample is an upsampling operation, which is used to increase the resolution of the feature map.

[0081] Step S3 is a process of locking the abnormal frequency domain based on the deep spectrum feature, capturing the abnormal texture description based on the deep texture feature, obtaining the abnormal smooth transition frame based on the deep time domain feature, and calculating the comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weight. Since the deep spectrum feature, the deep texture feature and the deep time domain feature are similar to the frequency domain feature, the texture feature and the time domain feature, only the accuracy is improved. Therefore, S31 locks the abnormal frequency domain based on the deep spectrum feature, including:

[0082] S311: performing spectrum analysis according to the deep spectrum features to form spectrum analysis data;

[0083] S312: decomposing the image information of the target sample into gradient divisions of different frequencies based on the spectrum analysis data, and forming a visible spectrum based on the gradient divisions;

[0084] S313: Locking a prominent area on the visible spectrum that is higher or lower than a preset spectrum threshold, and taking an area corresponding to the prominent area on the target sample as an abnormal area.

[0085] S32 captures abnormal texture description based on the deep texture feature, including:

[0086] S321: Acquire a unit pixel value and adjacent pixel values ​​of the unit pixel in the target sample based on the depth texture feature;

[0087] S322: Calculating a local binary pattern feature value based on the unit pixel value and the adjacent pixel values;

[0088] S323: Compare the local binary pattern feature value with a preset texture threshold to capture abnormal texture.

[0089] S33 acquires an abnormally smooth transition frame based on the deep temporal feature, including:

[0090] S31: Determine whether the depth time domain feature is empty; wherein,

[0091] If it is empty, output 0;

[0092] If it is not empty, inter-frame difference detection is performed according to the deep time domain features to obtain the pixel difference between adjacent frames, and the pixel difference is compared with a preset smoothing threshold. If the pixel difference is greater than the smoothing threshold, the two adjacent frames are regarded as abnormal smooth transition frames.

[0093] In step S3,

[0094] Calculating a comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and a preset module weight includes:

[0095] Calculate the frequency domain ratio of the abnormal frequency domain in the target sample, the area ratio of the abnormal texture description in the target sample, and the total frame ratio of the abnormal smooth transition frame in the target sample;

[0096] The frequency domain proportion is multiplied by a preset magnification value, and then multiplied by a preset frequency domain feature extraction module weight value to obtain a frequency domain score; the area proportion is multiplied by the magnification value, and then multiplied by a preset texture feature extraction module weight value to obtain a texture score; and the total frame proportion is multiplied by the magnification value, and then multiplied by a preset time domain feature extraction module weight value to obtain a time domain score;

[0097] The frequency domain score, the texture score, and the time domain score are added together to obtain a comprehensive score.

[0098] Then, step S4 is performed to compare the comprehensive score with a preset forgery threshold value to determine whether the target sample is a forged sample and form a determination result.

[0099] Step S5 is a process of performing reverse feedback based on the determination result to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of the neuron is lower than the preset usage threshold, the neuron is soft-pruned to update the enhanced FPN network. In this embodiment, the usage frequency of each neuron in the enhanced FPN network is displayed in the determination result, including:

[0100] The preset forward propagation network obtains the usage frequency and amplitude data of the neuron based on the preset neuron counter, and writes the usage frequency and amplitude data into the determination result; wherein,

[0101] The neuron counter is preset after each neuron in the enhanced FPN network. When the output of the neuron exceeds a preset output threshold, the neuron counter is incremented by one and the amplitude data of the neuron is recorded.

[0102] That is, neurons are monitored through a monitoring mechanism. The monitoring is implemented by adding a counter behind each neuron. Whenever the output of the neuron exceeds a certain threshold, the counter increases by 1. During training or testing, the activation frequency and amplitude data of each monitor are collected through the forward propagation network. After collecting a certain amount of data, the data of each monitor is analyzed to determine which neurons have a lower activation frequency or smaller amplitude. According to the analysis results, low-value neurons are soft pruned, thereby continuously optimizing the enhanced FPN network.

[0103] This soft pruning strategy can be seen as an adaptive optimization mechanism for the forgery detection model. As the model continues to learn new forgery features, the soft pruning strategy helps the model dynamically remove neurons that are no longer useful or have never been really useful, so as to reduce the complexity of the model and improve efficiency. Specifically, after each forgery detection task is completed, the model updates its internal state, including the activation frequency and amplitude of the neurons. These updated data are then used to decide which neurons can be pruned.

[0104] In the above process of real-time detection of target samples, the enhanced FPN network used each time adopts the latest updated enhanced FPN network to improve the detection quality.

[0105] As described above, the neural network-based counterfeit detection method provided by the present invention first performs feature extraction on the target sample through a preset frequency domain feature extraction module, a texture feature extraction module and a time domain feature extraction module to respectively obtain frequency domain features, texture features and time domain features, and then uses a preset enhanced FPN network to perform deep feature extraction on the frequency domain features, texture features and time domain features to obtain deep spectrum features, deep texture features and deep time domain features, and then locks the abnormal frequency domain based on the deep spectrum features, captures the abnormal texture description based on the deep texture features, obtains the abnormal smooth transition frame based on the deep time domain features, and calculates a comprehensive score based on the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weights, so as to determine whether the target sample is counterfeit, and performs reverse feedback based on the judgment result to obtain the usage frequency of each neuron in the enhanced FPN network, if the usage frequency of the neuron is lower than the preset usage threshold, the neuron is soft pruned to update the enhanced FPN network, so as to strengthen the enhanced FPN network during use, thereby continuously optimizing the enhanced FPN network, improving the accuracy and effectiveness of the enhanced FPN network, reducing the conditions for counterfeit detection, and reducing the difficulty of counterfeit detection.

[0106] like Figure 3 As shown, the present invention also provides a forgery detection system 100 based on a neural network, which implements the forgery detection method based on a neural network as described above, including a frequency domain feature extraction module 101, a texture feature extraction module 102, a time domain feature extraction module 103, an enhanced FPN network 104, a determination module 105, and a soft pruning module 106, wherein:

[0107] The frequency domain feature extraction module 101, the texture feature extraction module 102 and the time domain feature extraction module 103 are used to extract features from the target sample to obtain frequency domain features, texture features and time domain features respectively;

[0108] The enhanced FPN network 104 is used to perform deep feature extraction on the frequency domain features, the texture features and the time domain features to obtain deep spectrum features, deep texture features and deep time domain features;

[0109] The determination module 105 is used to lock the abnormal frequency domain based on the deep spectrum feature, capture the abnormal texture description based on the deep texture feature, obtain the abnormal smooth transition frame based on the deep time domain feature, and calculate the comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weight; and compare the comprehensive score with the preset forgery threshold to determine whether the target sample is a forged sample to form a determination result;

[0110] The soft pruning module 106 is used to perform reverse feedback based on the determination result to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of a neuron is lower than a preset usage threshold, the neuron is soft pruned to update the enhanced FPN network.

[0111] In this embodiment, the enhanced FPN network that performs deep feature extraction on the frequency domain features, the texture features and the time domain features in real time to obtain deep spectrum features, deep texture features and deep time domain features is the latest updated enhanced FPN network to improve the detection accuracy.

[0112] The specific implementation method is referred to the above-mentioned embodiment of the forgery detection method based on neural network, which will not be described in detail here.

[0113] As described above, the neural network-based counterfeit detection system provided by the present invention first performs feature extraction on the target sample through the preset frequency domain feature extraction module 101, texture feature extraction module 102 and time domain feature extraction module 103 to obtain frequency domain features, texture features and time domain features respectively, and then uses the preset enhanced FPN network 104 to perform deep feature extraction on the frequency domain features, texture features and time domain features to obtain deep spectrum features, deep texture features and deep time domain features, and then uses the judgment module 105 to lock the abnormal frequency domain based on the deep spectrum features, capture the abnormal texture description based on the deep texture features, and obtain the abnormal frequency domain based on the deep time domain features. Smooth transition frames, and calculate a comprehensive score based on the abnormal frequency domain, abnormal texture description, abnormal smooth transition frames and preset module weights to determine whether the target sample is forged, and perform reverse feedback based on the judgment result through the soft pruning module 106 to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of the neuron is lower than the preset usage threshold, the neuron is soft pruned to update the enhanced FPN network, thereby strengthening the enhanced FPN network during use, thereby continuously optimizing the enhanced FPN network, improving the accuracy and effectiveness of the enhanced FPN network, reducing the conditions for forgery detection, and reducing the difficulty of forgery detection.

[0114] As described above with reference to the accompanying drawings, the forgery detection method and system based on a neural network proposed in the present invention are described by way of example. However, those skilled in the art should understand that various improvements can be made to the forgery detection method and system based on a neural network proposed in the present invention without departing from the content of the present invention. Therefore, the protection scope of the present invention should be determined by the content of the attached claims.

Claims

1. A method for counterfeit detection based on neural network, characterized in that: include: The target sample is subjected to feature extraction by a preset frequency domain feature extraction module, a texture feature extraction module and a time domain feature extraction module to obtain frequency domain features, texture features and time domain features respectively; Using a preset enhanced FPN network to perform deep feature extraction on the frequency domain features, the texture features and the time domain features to obtain deep spectrum features, deep texture features and deep time domain features; Based on the deep spectrum feature, the abnormal frequency domain is locked, based on the deep texture feature, the abnormal texture description is captured, based on the deep time domain feature, the abnormal smooth transition frame is acquired, and a comprehensive score is calculated according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weight; Comparing the comprehensive score with a preset forgery threshold to determine whether the target sample is a forged sample and forming a determination result; Based on the determination result, reverse feedback is performed to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of a neuron is lower than a preset usage threshold, soft pruning is performed on the neuron to update the enhanced FPN network.

2. The neural network-based counterfeit detection method according to claim 1, characterized in that: Adopt the latest updated enhanced FPN network.

3. The neural network-based counterfeit detection method according to claim 2, characterized in that: The target sample is a target image or a target video; wherein, If the target sample is a target image, the time domain feature and the depth time domain feature are empty.

4. The neural network-based counterfeit detection method according to claim 3, characterized in that: The locking of the abnormal frequency domain based on the deep spectrum feature comprises: Performing spectrum analysis according to the deep spectrum features to form spectrum analysis data; Decomposing the image information of the target sample into gradient divisions of different frequencies based on the spectrum analysis data, and forming a visible spectrum based on the gradient divisions; A prominent area above or below a preset spectrum threshold is locked on the visible spectrum, and a region corresponding to the prominent area on the target sample is taken as an abnormal region.

5. The neural network-based counterfeit detection method according to claim 3, characterized in that: The capturing of abnormal texture description based on the deep texture feature comprises: Acquire a unit pixel value and adjacent pixel values ​​of the unit pixel in the target sample based on the depth texture feature; Calculate a local binary pattern feature value based on the unit pixel value and the adjacent pixel values; The local binary pattern feature value is compared with a preset texture threshold to capture abnormal texture.

6. The neural network-based counterfeit detection method according to claim 3, characterized in that: The acquiring of abnormally smooth transition frames based on the deep temporal features comprises: Determine whether the depth time domain feature is empty; wherein, If it is empty, output 0; If it is not empty, inter-frame difference detection is performed according to the deep time domain features to obtain the pixel difference between adjacent frames, and the pixel difference is compared with a preset smoothing threshold. If the pixel difference is greater than the smoothing threshold, the two adjacent frames are regarded as abnormal smooth transition frames.

7. The neural network-based counterfeit detection method according to any one of claims 4 to 6, characterized in that: Calculating a comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and a preset module weight includes: Calculate the frequency domain ratio of the abnormal frequency domain in the target sample, the area ratio of the abnormal texture description in the target sample, and the total frame ratio of the abnormal smooth transition frame in the target sample; The frequency domain proportion is multiplied by a preset magnification value, and then multiplied by a preset frequency domain feature extraction module weight value to obtain a frequency domain score; the area proportion is multiplied by the magnification value, and then multiplied by a preset texture feature extraction module weight value to obtain a texture score; and the total frame proportion is multiplied by the magnification value, and then multiplied by a preset time domain feature extraction module weight value to obtain a time domain score; The frequency domain score, the texture score, and the time domain score are added together to obtain a comprehensive score.

8. The neural network-based counterfeit detection method according to claim 1, characterized in that: The determination result shows the usage frequency of each neuron in the enhanced FPN network, including: The preset forward propagation network obtains the usage frequency and amplitude data of the neuron based on the preset neuron counter, and writes the usage frequency and amplitude data into the determination result; wherein, The neuron counter is preset after each neuron in the enhanced FPN network. When the output of the neuron exceeds a preset output threshold, the neuron counter is incremented by one and the amplitude data of the neuron is recorded.

9. A counterfeit detection system based on a neural network, characterized in that: To implement the neural network-based counterfeit detection method as described in any one of claims 1 to 8, the system includes a frequency domain feature extraction module, a texture feature extraction module, a time domain feature extraction module, an enhanced FPN network, a determination module, and a soft pruning module, wherein: The frequency domain feature extraction module, the texture feature extraction module and the time domain feature extraction module are used to respectively extract features from the target sample to obtain frequency domain features, texture features and time domain features; The enhanced FPN network is used to perform deep feature extraction on the frequency domain features, the texture features and the time domain features to obtain deep spectrum features, deep texture features and deep time domain features; The determination module is used to lock the abnormal frequency domain based on the deep spectrum feature, capture the abnormal texture description based on the deep texture feature, obtain the abnormal smooth transition frame based on the deep time domain feature, and calculate the comprehensive score according to the abnormal frequency domain, the abnormal texture description, the abnormal smooth transition frame and the preset module weight; and compare the comprehensive score with the preset forgery threshold to determine whether the target sample is a forged sample to form a determination result; The soft pruning module is used to perform reverse feedback based on the determination result to obtain the usage frequency of each neuron in the enhanced FPN network. If the usage frequency of a neuron is lower than a preset usage threshold, the neuron is soft pruned to update the enhanced FPN network.

10. The neural network-based counterfeit detection system according to claim 9, characterized in that: The enhanced FPN network that performs deep feature extraction on the frequency domain features, the texture features and the time domain features in real time to obtain deep spectrum features, deep texture features and deep time domain features is the latest updated enhanced FPN network.

Images