WO2023071180A1 - 真伪识别方法、装置、电子设备以及存储介质 - Google Patents

真伪识别方法、装置、电子设备以及存储介质 Download PDF

Info

Publication number
WO2023071180A1
WO2023071180A1 PCT/CN2022/096019 CN2022096019W WO2023071180A1 WO 2023071180 A1 WO2023071180 A1 WO 2023071180A1 CN 2022096019 W CN2022096019 W CN 2022096019W WO 2023071180 A1 WO2023071180 A1 WO 2023071180A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
preset
dimension
identified
feature
Prior art date
Application number
PCT/CN2022/096019
Other languages
English (en)
French (fr)
Inventor
梁嘉豪
石华峰
吴一超
梁鼎
Original Assignee
上海商汤智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海商汤智能科技有限公司 filed Critical 上海商汤智能科技有限公司
Publication of WO2023071180A1 publication Critical patent/WO2023071180A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present disclosure relates to the technical field of image processing, and in particular to an authenticity identification method, device, electronic equipment and storage medium.
  • Embodiments of the present disclosure at least provide an authenticity identification method, device, electronic equipment, and storage medium.
  • an embodiment of the present disclosure provides a method for authenticity identification, which is applied to a target deep neural network, including:
  • At least one of a forged region in the image to be identified and authenticity result information of the object to be identified is determined.
  • using the target deep neural network can more accurately determine the image features describing multiple preset dimensions of the image to be recognized, and then use the determined image features to more accurately identify whether the object to be recognized is a forged object, that is, to obtain a more accurate Accurate true and false result information, and more accurate forged areas can be obtained.
  • these two detection tasks can promote each other and share some specific image features, which can not only improve detection efficiency, but also enhance The ability to extract image features improves the detection accuracy of two detection tasks.
  • the determining at least one of the forged region in the image to be identified and the authenticity result information of the object to be identified based on the image features includes:
  • the number of feature points in the first image feature of the highest preset dimension is small, and the feature data dimension corresponding to a feature point is relatively high, thus effectively removing redundant information in the image to be recognized, while retaining and increasing Effective information for authenticity identification, so it can effectively improve the accuracy of the determined authenticity result information;
  • the number of feature points in the second image feature corresponding to the lowest preset dimension is large, so the corresponding image to be identified can be determined more accurately The authenticity information corresponding to each pixel in the image, and then can more accurately determine the forged area in the image to be recognized.
  • the feature data dimension corresponding to each feature point in the second image feature corresponding to the lowest preset dimension is low , so the speed of determining the forged region can be improved.
  • the image features include a second image feature corresponding to each preset dimension in multiple preset dimensions, and a first image feature corresponding to each preset dimension in multiple preset dimensions ;
  • the extraction of image features corresponding to multiple preset dimensions of the image to be recognized includes:
  • performing operations such as dimensionality reduction and fusion on the first image features of each preset dimension can more accurately determine that the image to be recognized corresponds to the second image feature of each preset dimension in multiple preset dimensions.
  • the determining, based on the first image features corresponding to each preset dimension, that the image to be recognized corresponds to a second image feature of each preset dimension in multiple preset dimensions includes:
  • the first dimension group includes a first preset dimension and a second preset dimension, and the first preset dimension is higher than the second preset dimension;
  • a first feature processing operation is performed on the second image feature corresponding to the first preset dimension in the first dimension group, to obtain the same as the second preset A third image feature whose dimensions match; wherein, the feature map corresponding to the third image feature has the same image resolution as the feature map corresponding to the first image feature of the second preset dimension;
  • the first image feature corresponding to the higher preset dimension among the two adjacent preset dimensions is processed, and the processed third image feature is the same as the first image feature corresponding to the lower preset dimension.
  • the corresponding feature maps have the same image resolution; subsequent feature fusion based on the third image feature and the first image feature with the same data dimension and image resolution can improve the fusion accuracy and obtain a more accurate second image features.
  • the first image feature corresponding to the second preset dimension in the first dimension group is determined based on the obtained third image feature and the first dimension group
  • the second image features corresponding to the second preset dimension include:
  • a second image feature corresponding to a second preset dimension in the first dimension group is determined.
  • the above-mentioned third image feature and the first image feature specifically have the same data dimension and the same image resolution, so the two can be accurately spliced; after that, feature extraction and Processing and other operations can obtain more accurate processing results, that is, the above-mentioned second image features.
  • the extracting the first image feature of the image to be recognized corresponding to each preset dimension in multiple preset dimensions includes:
  • the second dimension group includes a third preset dimension and a fourth preset dimension, and the third preset dimension is lower than said fourth preset dimension;
  • a second feature processing operation is performed on the first image feature corresponding to the third preset dimension in the second dimension group to obtain the same as the fourth preset Dimensionally matched fourth image features;
  • the first image feature is processed according to the fourth preset dimension, and the fourth image feature that matches the fourth preset dimension can be determined more accurately; the fourth image feature is then continuously processed to obtain The first image feature of matches the fourth preset dimension.
  • the first image features of lower preset dimensions in each second dimension group are sequentially processed, so that the first image features corresponding to each preset dimension can be determined more accurately.
  • the obtaining the authenticity result information of the object to be identified based on the first image feature corresponding to the highest preset dimension among the plurality of preset dimensions includes:
  • the feature data corresponding to each feature point has more dimensions, and the number of feature points is less. Image features. Therefore, using the first image features, the above-mentioned first predicted probability and second predicted probability can be determined more accurately, and then, based on the first predicted probability and second predicted probability, more accurate authenticity result information can be obtained.
  • the determining the forged region in the image to be identified based on the second image feature corresponding to the lowest preset dimension among the plurality of preset dimensions includes:
  • a forged area in the image to be identified is determined.
  • the second image feature corresponding to the lowest preset dimension is obtained by reducing the dimensionality, adding corresponding feature points, and concatenating the first image features corresponding to the preset dimensions of each level. Therefore, each of the second image features
  • the feature data corresponding to each feature point can more accurately characterize whether the corresponding feature point is a forged feature point; It can be more accurately determined that each pixel point in the image to be recognized is a forged pixel point forgery result information, and then a more accurate forgery area can be determined.
  • the acquisition includes an image to be recognized of the object to be recognized, including:
  • an image region corresponding to the object to be recognized is extracted from the original image to obtain the image to be recognized.
  • the detection frame and the key points can respectively determine the image area occupied by the object to be recognized in the original image, and the combination of the two to determine the above image area can play a role in mutual calibration, so it can be more accurate. image area, that is, a more accurate image to be recognized can be obtained.
  • the extracting the image region corresponding to the object to be recognized from the original image based on the detection frame and the plurality of key points to obtain the image to be recognized includes:
  • an image area corresponding to the object to be identified is extracted from the original image to obtain the image to be identified.
  • the area corresponding to the target area information includes a complete object to be identified, which is beneficial to improving the accuracy of authenticity identification corresponding to the object to be identified.
  • the extracting the image region corresponding to the object to be recognized from the original image based on the detection frame and the plurality of key points to obtain the image to be recognized includes:
  • the image area corresponding to the object to be identified is extracted from the original image, and the obtained The image to be recognized.
  • the image region corresponding to the object to be recognized that occupies a larger area occupied by the original image is extracted, which can ensure a larger resolution of the obtained image to be recognized, which is beneficial to improve the accuracy of authenticity recognition.
  • the above method further includes:
  • a heat map is generated based on the forged area and the image to be identified; wherein, the heat value of a pixel point in the heat map corresponding to the forged area is higher than a preset value.
  • the thermal map is used to realize the visualization of the forged area corresponding to the image to be recognized, which improves the intuitiveness of the forged area.
  • the above-mentioned authenticity identification method also includes the step of training the target deep neural network:
  • the sample image is input into the target deep neural network to be trained, and the multiple sample images are processed through the target neural network to obtain the first prediction score, the The second predicted score of the sample object being a forged object, and the predicted probability information that each pixel in each sample image is a forged pixel;
  • the target deep neural network to be trained is trained by using the network loss information until the preset training condition is met, and a trained target deep neural network is obtained.
  • the prediction result of the forged region can be directly obtained, so the above prediction probability information can be used to characterize the prediction result of the forgery region; through the first prediction score and the second prediction score
  • the value can determine the authenticity identification result of the sample object. Therefore, based on the first predicted score and the second predicted score (corresponding to the detection task of authenticity identification), and the predicted probability information and the standard probability information (corresponding to the detection task of the forged region), these two detection tasks
  • the prediction value is used to establish the network loss information of the training target neural network, which can effectively improve the detection accuracy of the trained target neural network through the mutual promotion of the two detection tasks.
  • the generating network loss information based on the first prediction score, the second prediction score, prediction probability information and standard probability information corresponding to each sample image includes:
  • the network loss information is generated based on the first loss information and the second loss information.
  • the above-mentioned first loss information can be determined more accurately by using the authenticity identification information of the sample image, that is, the above-mentioned first prediction score and the second prediction score; Probability information) and standard results (standard probability information), the above second loss information can be determined more accurately; then, based on the first loss information and the second loss information, network loss information representing the losses of the two detection tasks can be generated.
  • an authenticity identification device including:
  • An image acquisition module configured to acquire an image to be identified including an object to be identified
  • a feature extraction module configured to extract a first image feature corresponding to a plurality of preset dimensions of the image to be identified; wherein, the number of feature points corresponding to the first image feature is negatively correlated with the value of the corresponding preset dimension;
  • a detection module configured to determine at least one of a forged region in the image to be identified and authenticity result information of the object to be identified based on the image features.
  • an embodiment of the present disclosure further provides an electronic device, including: a processor, a memory, and a bus, the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processing
  • the processor communicates with the memory through a bus, and when the machine-readable instructions are executed by the processor, the above-mentioned first aspect, or the steps in any possible implementation manner of the first aspect are executed.
  • embodiments of the present disclosure further provide a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned first aspect, or any of the first aspects of the first aspect, may be executed. Steps in one possible implementation.
  • FIG. 1 shows a flow chart of a method for authenticity identification provided by an embodiment of the present disclosure
  • FIG. 2 shows a flow chart of another authenticity identification method provided by an embodiment of the present disclosure
  • FIG. 3 shows a flowchart of a network training method provided by an embodiment of the present disclosure
  • Fig. 4 shows a schematic diagram of an authenticity identification device provided by an embodiment of the present disclosure
  • Fig. 5 shows a schematic diagram of an electronic device provided by an embodiment of the present disclosure.
  • the target deep neural network can be used to accurately determine the image features describing multiple preset dimensions of the image to be identified, and then use the determined image features to accurately identify whether the object to be identified is a fake object, that is, to obtain a more accurate The true and false result information, and the more accurate forged area can be obtained.
  • these two detection tasks can promote each other and share some specific image features, which can not only improve detection efficiency, but also enhance The ability to extract image features can extract the key information corresponding to authenticity identification in the image, and improve the detection accuracy of the two detection tasks.
  • the authenticity identification method provided by the embodiment of the present disclosure will be described below by taking the executing subject as a device capable of computing as an example.
  • the authenticity identification method provided by the present disclosure is applied to the target deep neural network, and may include the following steps:
  • the object to be identified may be an object that needs to be authenticated, for example, a human face, and the object to be identified may be determined according to a specific application scenario, which is not limited in the present disclosure.
  • the above-mentioned image to be recognized may be taken by the above-mentioned device with computing capability, or it may be taken by other shooting devices and transmitted to the above-mentioned device with computing capability, which is not limited in the present disclosure.
  • the above image to be recognized may be a captured original image, or may be a sub-image corresponding to the object to be recognized intercepted from the captured original image, which is not limited in the present disclosure.
  • the aforementioned image to be recognized may be selected from a video clip, or may be an independently existing image, which is not limited in the present disclosure.
  • the aforementioned image features may include a second image feature corresponding to each preset dimension in the plurality of preset dimensions, and a first image feature corresponding to each preset dimension in the plurality of preset dimensions.
  • the above-mentioned image features can be extracted by using the following steps: first extract the first image feature corresponding to each preset dimension in the multiple preset dimensions of the image to be recognized, and then, based on the first image feature corresponding to each preset dimension Image features, determining that the image to be recognized corresponds to a second image feature of each preset dimension in a plurality of preset dimensions.
  • Different preset dimensions correspond to different numbers of feature points of the first image feature, and data dimensions of feature data corresponding to the feature points are also different.
  • the higher the preset dimension the smaller the number of corresponding feature points of the first image feature, and the higher the data dimension of feature data corresponding to a feature point, the smaller the corresponding feature map of the first image feature.
  • the preset dimensions are pre-set according to specific application scenarios.
  • feature extraction is first performed on the image to be recognized, and the original image features corresponding to the lowest preset dimension are obtained, and then the extracted image features are processed by using a convolution layer, etc., to obtain the first image feature corresponding to the lowest preset dimension .
  • the adjacent first image features corresponding to the higher preset dimensions are obtained.
  • the feature map corresponding to the second image feature has the same size and resolution as the image to be recognized.
  • the second image feature corresponding to the lowest preset dimension is finally determined by performing operations such as reducing the data dimension of the feature data and concatenating the first image feature in descending order of the preset dimensions.
  • the feature data corresponding to each feature point of the second image feature can more accurately represent whether the corresponding feature point is a forged feature point.
  • the authenticity result information of the object to be identified can be obtained based on the first image feature corresponding to the highest preset dimension among the plurality of preset dimensions; and/or, based on one of the plurality of preset dimensions
  • the second image feature corresponding to the lowest preset dimension determines the forged region in the image to be identified.
  • the feature data corresponding to each feature point in the first image feature corresponding to the highest preset dimension has more dimensions and fewer feature points.
  • image features can more accurately describe image features that are effective for authenticity identification, so The authenticity result information of the object to be identified can be determined more accurately by using the first image feature.
  • a convolutional layer is used to process the first image feature corresponding to the highest preset dimension to obtain the authenticity result information.
  • the second image feature corresponding to the lowest preset dimension is the first image feature corresponding to each level of preset dimension through dimensionality reduction, adding corresponding feature points, and splicing operations. Therefore, each feature point of the second image feature corresponds to The feature data can more accurately characterize whether the corresponding feature point is a forged feature point; and because the feature map corresponding to the second image feature has the same resolution as the image to be recognized, therefore, using the second image feature can be more accurately It is determined that each pixel in the image to be recognized is the forgery result information of a forged pixel, and then a more accurate forged area can be determined.
  • a convolutional layer is used to process the second image feature corresponding to the aforementioned lowest preset dimension to obtain the aforementioned forged region.
  • the following steps can be used to determine the second image features corresponding to each preset dimension:
  • group the preset dimensions take every two adjacent preset dimensions as a group to obtain multiple first dimension groups; the first dimension groups include the first preset dimension and the second preset dimension, The first preset dimension is higher than the second preset dimension.
  • the dimension of a certain first dimension group may be the first preset dimension in the first dimension group.
  • the second image corresponding to the first preset dimension in the first dimension group Perform a feature processing operation on the feature to obtain a third image feature that matches the second preset dimension; wherein, the feature map corresponding to the third image feature corresponds to the first image feature corresponding to the second preset dimension
  • the feature maps of have the same image resolution and size. Based on the obtained third image feature and the first image feature corresponding to the second preset dimension in the first dimension group, determine the second image feature corresponding to the second preset dimension in the first dimension group .
  • the feature map corresponding to the fifth image feature has the same image resolution and size as the feature map corresponding to the first image feature corresponding to the second preset dimension. Based on the obtained fifth image feature and the first image feature corresponding to the second preset dimension in the first dimension group, determine the second image feature corresponding to the second preset dimension in the first dimension group .
  • the above-mentioned feature processing operation may be a transposed convolution operation, through which the data dimension of the feature data corresponding to the feature points can be reduced while increasing the number of feature points, that is, the resolution of the corresponding feature map can be improved.
  • the first image feature 291 corresponding to the higher preset dimension among the two adjacent preset dimensions in the first dimension group is subjected to a feature processing operation, and the processing
  • the last fifth image feature has the same data dimension as the first image feature corresponding to the lower preset dimension, and the corresponding feature map has the same image resolution; after that, the feature is processed into the third image feature after operation
  • the first image feature corresponding to the second preset dimension in the first dimension group is spliced to obtain a spliced image feature, such as the spliced image feature 21 shown in FIG. 2 .
  • the second image features 23 corresponding to the second preset dimensions in the first dimension group are determined. For example, at least one convolution process may be performed on the spliced image features 21 to obtain the second image features 23.
  • a stitched image feature 21 , a second image feature 23 , etc. are shown in FIG. 2 .
  • the second image feature 23 corresponding to the higher preset dimension among the two adjacent preset dimensions in the first dimension group Perform a feature processing operation, the processed third image feature has the same data dimension as the first image feature corresponding to the lower preset dimension, and the corresponding feature map has the same image resolution; after that, the third image
  • the feature is concatenated with the first image feature corresponding to the second preset dimension in the first dimension group to obtain the concatenated image feature, as shown in FIG. 2 , the concatenated image feature 22 .
  • the second image features 24 corresponding to the second preset dimensions in the first dimension group are determined. For example, at least one convolution process may be performed on the spliced image features 22 to obtain the second image features 24.
  • the above-mentioned third image feature corresponding to the same first dimension group and the first image feature corresponding to the second preset dimension have the same data dimension and the same image resolution, so the two can be spliced accurately; after that, based on Operations such as feature extraction and processing are performed on the spliced image features after splicing, and relatively accurate processing results can be obtained, that is, the above-mentioned second image features.
  • the first image feature and the second image feature in each first dimension group are sequentially processed in order of dimensions from high to low, and the second image feature corresponding to the lowest preset dimension can be obtained.
  • the following steps can be used to determine the forged region in the image to be identified:
  • the probability information that each pixel in the image to be recognized is a fake pixel is determined.
  • the second image feature 25 of the lowest preset dimension is shown in FIG. 2 .
  • the feature map corresponding to the second image feature of the lowest preset dimension has the same resolution and size as the image to be recognized. Therefore, each pixel in the feature map corresponding to the second image feature is identical to each pixel in the image to be recognized. If the pixel points correspond to each other, then, using the second image feature, it can be more accurately determined that each pixel point in the image to be recognized is a forged pixel point forgery result information.
  • the second image feature 25 corresponding to the lowest preset dimension can be processed by using a fully connected network layer, a classifier, etc., to obtain probability information that each pixel in the image to be recognized is a fake pixel.
  • each pixel in the image to be recognized After obtaining the probability information corresponding to each pixel in the image to be recognized, based on the determined probability information and a preset probability threshold, determine forgery result information that each pixel in the image to be recognized is a forged pixel. Based on the forgery result information corresponding to each pixel, a forged area in the image to be identified is determined.
  • the aforementioned preset probability thresholds are flexibly set according to specific application scenarios.
  • the probability value corresponding to the above probability information is greater than or equal to the above preset probability threshold, it is determined that the pixel corresponding to the above probability information is a fake pixel; when the probability value corresponding to the above probability information is less than the above preset probability threshold , and determine that the pixel corresponding to the above probability information is a pixel that has not been tampered with.
  • the pixel points determined to be fake pixel points may form at least one fake region.
  • a mask map M pred with the same size as the image to be identified is created. Afterwards, the MASK map M pred is filled according to the following formula:
  • (i, j) represents the row and column identifier of the corresponding pixel
  • ⁇ represents the above-mentioned preset probability threshold
  • I pred represents the probability value corresponding to the above-mentioned probability information.
  • a heat map may be generated based on the forged area and the image to be identified; wherein, the heat value of a pixel point in the heat map corresponding to the forged area is higher than a preset value.
  • the resolution and size of the heat map are determined according to the resolution and size of the image to be recognized, for example, the heat map may be set to have the same size and resolution as the image to be recognized. Afterwards, the heat values of the pixels corresponding to the forged area in the heat map can be higher than the preset value, and the heat values of the pixels in the heat map are equal;
  • the above probability information sets the heat value of the corresponding pixel in the heat map. Specifically, as the probability value corresponding to the probability information increases, the heat value of the corresponding pixel also increases.
  • the heat values corresponding to pixels outside the forged area may be set to be equal, or set according to the probability value corresponding to the above probability information, which is not limited in the present disclosure.
  • the visualization of the forged area corresponding to the image to be recognized is realized by using the heat map, which improves the intuition and interpretability of the forged area.
  • the following steps can be used to extract the first image feature of the image to be recognized corresponding to each preset dimension in multiple preset dimensions:
  • the image to be recognized can be input into the target deep neural network, and then undergo at least one column-depth separable convolution operation to obtain the first image with the lowest preset dimension.
  • Feature 26 For example, the image to be recognized can be input into the target deep neural network, and then undergo at least one column-depth separable convolution operation to obtain the first image with the lowest preset dimension.
  • every two adjacent preset dimensions can be used as a group to obtain a plurality of second dimension groups; the second dimension group includes the third preset dimension and the second dimension Four preset dimensions, the third preset dimension is lower than the fourth preset dimension.
  • the dimension of the second dimension group may be the third preset dimension in the second dimension group.
  • each second dimension group After the above-mentioned second dimension group is determined, perform the following operations on each second dimension group in order of dimensions from low to high until the first image feature of each preset dimension except the lowest preset dimension is determined:
  • a second feature processing operation is performed on the first image feature corresponding to the third preset dimension in the second dimension group to obtain the same as the fourth preset A fourth image feature whose dimensions match; based on the obtained fourth image feature, determine a first image feature corresponding to a fourth preset dimension in the second dimension group.
  • the first image feature 26 corresponding to the lower preset dimension in the second dimension group is subjected to a feature processing operation to obtain a fourth image matching the higher preset dimension feature 29, and then, based on the obtained fourth image feature 29, after operations such as convolution, determine the first image feature 27 corresponding to a higher preset dimension.
  • the above-mentioned second feature processing operation may be a separable convolution operation, which is specifically used to increase the data dimension of the feature data of the feature points in the first image feature and reduce the number of feature points.
  • the fourth image feature obtained after the second feature processing operation matches the higher fourth preset dimension in the second dimension group.
  • At least one convolution operation can be performed to obtain the first image feature corresponding to the higher preset dimension in the second dimension group , the first image feature and the fourth image feature have the same preset dimension, and the corresponding feature maps have the same image resolution.
  • the first image feature in the corresponding second dimension group is processed, and the fourth image feature matching the fourth preset dimension can be determined more accurately;
  • the fourth image feature is further processed, and the obtained first image feature matches the fourth preset dimension.
  • the first image features of lower preset dimensions in each second dimension group are sequentially processed, so that the first image features corresponding to each preset dimension can be determined more accurately.
  • FIG. 2 shows the first image feature 27, the first image feature 28, etc. in the second dimension group, and the fourth image feature 29, etc. in the second dimension group.
  • the following steps can be used to determine the authenticity result information of the object to be identified:
  • First determine two scores specifically, based on the first image feature corresponding to the highest preset dimension, determine the first score for the object to be recognized as a real object and the second score for the object to be recognized as a counterfeit object .
  • the first image feature may be processed through at least one convolution operation to obtain a first score indicating that the object to be identified is a real object and a second score indicating that the object to be identified is a fake object.
  • determine two predicted probabilities specifically, based on the first score and the second score, determine the first predicted probability that the object to be identified is a real object and the object to be identified is The second predicted probability of a fake object.
  • i and class both represent the identity of the object to be recognized as a real object or a fake object
  • i is 0, which means the object to be recognized is a real object
  • i is 1, which means the object to be recognized is a fake object
  • p represents the prediction probability
  • class is 0
  • p class represents the first predicted probability that the object to be recognized is a real object
  • p class represents the second predicted probability that the object to be recognized is a forged object
  • x represents the score
  • x[i ] represents the first score of the object to be recognized as a real object
  • x[i] represents the second score of the object to be recognized as a forged object.
  • a classifier may be used to determine the above two prediction probabilities.
  • the authenticity result prediction is performed, specifically, the authenticity result information of the object to be identified is determined based on the first prediction probability and the second prediction probability.
  • the first predicted probability can be compared with the second predicted probability, and the identification result corresponding to the larger predicted probability can be used as the authenticity result information. For example, if the first predicted probability is greater than the second predicted probability, then The authenticity result information indicates that the object to be identified is a real object, and if the first predicted probability is less than or equal to the second predicted probability, the authenticity result information indicates that the object to be identified is a forged object.
  • c is a parameter used to characterize the authenticity result information.
  • the feature data corresponding to each feature point in the first image feature corresponding to the highest preset dimension has more dimensions, and the number of feature points is less.
  • image features can be described more accurately and are effective for authenticity identification. Therefore, using the first image feature, the above-mentioned first and second prediction probabilities can be determined more accurately, and then, based on the first and second prediction probabilities, more accurate authenticity result information can be obtained.
  • the image to be recognized in the above embodiments may be directly captured by the photographing device, or may be a sub-image intercepted from the captured original image.
  • the image to be recognized corresponding to the object to be recognized can be intercepted by the following steps:
  • the above-mentioned original image may be obtained from a video clip.
  • frame extraction processing may be performed at equal intervals from a video clip to obtain multiple frames of original images, and then implement the disclosed method on each frame of the original image. Realize the authenticity identification of each frame image.
  • Both the detection frame and the key points can respectively determine the image area occupied by the object to be recognized in the original image. Combining the two to determine the above image area can play a role in mutual calibration, so a more accurate image area can be obtained. That is, a more accurate image to be recognized can be obtained.
  • the number of key points located in the detection frame is counted first; then, based on the counted quantity, the proportion of the key points located in the detection frame in all key points is determined, and when the proportion is greater than the preset proportion
  • the above image region can be determined based on the position of the key point and the position of the detection frame.
  • the determined image area may be only the area corresponding to the detection frame, or may include the area corresponding to the detection frame and all key points.
  • the object recognition can be re-determined, and the detection frame and key points of the object to be recognized can be re-determined.
  • the image region corresponding to the object to be recognized is extracted from the original image to obtain the image to be recognized. Specifically, the following steps can be used to achieve :
  • the area corresponding to the above initial area information may be an area including a detection frame and a preset number of key points.
  • the area corresponding to the target area information includes the complete object to be identified and a small part of the environment around the object, which is conducive to improving the accuracy of authenticity identification corresponding to the object to be identified.
  • the image to be recognized when the image to be recognized is recognized, not only the detection frame and key points can be obtained, but also the area of the area occupied by the object to be recognized in the original image (hereinafter referred to as area area information) and the area of the object to be recognized can be determined.
  • area area information which can be saved in a json file for subsequent processing.
  • the required information can be extracted from the json file.
  • the area area information can be obtained from the json file first, and in the case that the area area corresponding to the area area information is larger than the preset area, the detection frame and the multiple key points, extract the image area corresponding to the object to be recognized from the original image, and obtain the image to be recognized.
  • the obtained image to be recognized can be saved as a picture in png format.
  • Extracting the image region corresponding to the object to be recognized that occupies a larger area in the original image can ensure a larger resolution of the obtained image to be recognized, which is conducive to improving the accuracy of authenticity recognition.
  • the present disclosure also provides a training method of a target deep neural network, as shown in FIG. 3 , which may include the following steps:
  • the sample image above is an image including a sample object, for example, a sample image including a human face.
  • the sample image here may be the original image taken by the shooting device, or a sub-image intercepted from the original image including the object to be recognized.
  • the above network loss information includes the first loss information corresponding to the authenticity detection of the sample object and the second loss information corresponding to the determined loss area.
  • the first sample probability that the sample object is a real object and the second sample probability that the sample object is a fake object can be determined.
  • the calculation methods of the first sample probability and the second sample probability are the same as the above-mentioned calculation methods of the first prediction probability and the second prediction probability, and will not be repeated here.
  • the first loss information can be generated using the following formula:
  • L c represents the first loss information
  • i represents the identity of the sample object as a real object or a fake object
  • i is 0, which means the sample object is a real object
  • i is 1, which means the sample is a fake object
  • p represents the sample probability
  • p 0 Indicates the first sample probability that the sample object is a real object
  • p 1 represents the second sample probability that the sample object is a fake object
  • q represents the standard probability
  • q 0 represents the first standard probability that the sample object is a real object
  • q 1 represents the sample The second standard probability that the object is a fake object.
  • the second loss information can be generated using the following formula:
  • L region represents the second loss information
  • (i, j) represents the row and column identifier of the corresponding pixel
  • I pred represents the predicted probability information corresponding to the corresponding pixel
  • Mtarget represents the standard probability information corresponding to the corresponding pixel.
  • the network loss information may be generated based on the first loss information and the second loss information.
  • the following formula can be used to realize:
  • L represents the network loss information
  • a and b represent the preset weights.
  • the prediction result of the forged region can be directly obtained, so the above prediction probability information can be used to characterize the prediction result of the forgery region; through the first prediction score and the second prediction
  • the score can determine the authenticity identification result of the sample object. Therefore, based on the first predicted score and the second predicted score (corresponding to the detection task of authenticity identification), and the predicted probability information and the standard probability information (corresponding to the detection task of the forged region), these two detection tasks
  • the prediction value is used to establish the network loss information of the training target neural network, which can effectively improve the detection accuracy of the trained target neural network through the mutual promotion of the two detection tasks.
  • the writing order of each step does not mean a strict execution order and constitutes any limitation on the implementation process.
  • the specific execution order of each step should be based on its function and possible
  • the inner logic is OK.
  • the embodiment of the present disclosure also provides an authenticity identification device corresponding to the authenticity identification method. Since the problem-solving principle of the device in the embodiment of the present disclosure is similar to the above-mentioned authenticity identification method in the embodiment of the disclosure, the device For the implementation, please refer to the implementation of the method, and the repeated parts will not be repeated.
  • FIG. 4 it is a schematic diagram of the structure of an authenticity identification device provided by an embodiment of the present disclosure, and the device includes:
  • An image acquiring module 410 configured to acquire an image to be recognized including an object to be recognized.
  • the feature extraction module 420 is configured to extract image features corresponding to multiple preset dimensions of the image to be recognized; wherein, the number of feature points corresponding to the image feature is negatively correlated with the value of the corresponding preset dimension.
  • the detection module 430 is configured to determine at least one of a forged region in the image to be recognized and authenticity result information of the object to be recognized based on the image feature.
  • the detection module 430 is specifically configured to:
  • the image feature includes a second image feature corresponding to each preset dimension in a plurality of preset dimensions, and a first image feature corresponding to each preset dimension in a plurality of preset dimensions;
  • the feature extraction module 420 extracts image features corresponding to multiple preset dimensions of the image to be recognized, it is used to:
  • the feature extraction module 420 determines that the image to be recognized corresponds to the second image feature of each preset dimension in multiple preset dimensions based on the first image feature corresponding to each preset dimension , specifically for:
  • the first dimension group includes a first preset dimension and a second preset dimension, and the first preset dimension is higher than the second preset dimension;
  • the first feature processing operation is performed on the first image feature corresponding to the first preset dimension in the first dimension group, to obtain the second preset dimension.
  • a third image feature whose dimensions match; wherein, the feature map corresponding to the third image feature has the same image resolution as the feature map corresponding to the first image feature of the second preset dimension;
  • the feature extraction module 420 determines the first dimension group based on the obtained third image feature and the first image feature corresponding to the second preset dimension in the first dimension group When the second image feature corresponding to the second preset dimension in is used for:
  • a second image feature corresponding to a second preset dimension in the first dimension group is determined.
  • the feature extraction module 420 when the feature extraction module 420 extracts the first image feature of the image to be recognized corresponding to each preset dimension in multiple preset dimensions, it is used to:
  • the second dimension group includes a third preset dimension and a fourth preset dimension, and the third preset dimension is lower than said fourth preset dimension;
  • a second feature processing operation is performed on the first image feature corresponding to the third preset dimension in the second dimension group to obtain the same as the fourth preset Dimensionally matched fourth image features;
  • the detection module 430 is configured to: when obtaining the authenticity result information of the object to be identified based on the first image feature corresponding to the highest preset dimension among the multiple preset dimensions:
  • the detection module 430 is configured to:
  • a forged area in the image to be identified is determined.
  • the image acquisition module 410 when the image acquisition module 410 acquires the image to be identified including the object to be identified, it is used to:
  • an image region corresponding to the object to be recognized is extracted from the original image to obtain the image to be recognized.
  • the image acquisition module 410 extracts the image region corresponding to the object to be recognized from the original image based on the detection frame and the plurality of key points, and obtains the image to be recognized , for:
  • an image area corresponding to the object to be identified is extracted from the original image to obtain the image to be identified.
  • the image acquisition module 410 extracts the image region corresponding to the object to be recognized from the original image based on the detection frame and the plurality of key points, and obtains the image to be recognized , for:
  • the image area corresponding to the object to be identified is extracted from the original image, and the obtained The image to be recognized.
  • the detection module 430 is further configured to:
  • a heat map is generated based on the forged area and the image to be identified; wherein, the heat value of a pixel point in the heat map corresponding to the forged area is higher than a preset value.
  • the above-mentioned device also includes a training module 440 for training the target deep neural network, and the training module 440 is used for:
  • the sample image is input into the target deep neural network to be trained, and the multiple sample images are processed through the target neural network to obtain the first prediction score, the The second predicted score of the sample object being a forged object, and the predicted probability information that each pixel in each sample image is a forged pixel;
  • the target deep neural network to be trained is trained by using the network loss information until the preset training condition is met, and a trained target deep neural network is obtained.
  • the training module 440 is used to:
  • the network loss information is generated based on the first loss information and the second loss information.
  • an embodiment of the present disclosure also provides an electronic device.
  • FIG. 5 it is a schematic structural diagram of an electronic device 500 provided by an embodiment of the present disclosure, including a processor 51 , a memory 52 , and a bus 53 .
  • the memory 52 is used to store execution instructions, including a memory 521 and an external memory 522; the memory 521 here is also called an internal memory, and is used to temporarily store the calculation data in the processor 51 and the data exchanged with the external memory 522 such as a hard disk,
  • the processor 51 exchanges data with the external memory 522 through the memory 521.
  • the processor 51 communicates with the memory 52 through the bus 53, so that the processor 51 executes the following instructions:
  • Embodiments of the present disclosure also provide a computer-readable storage medium, on which a computer program is stored, and when the computer program is run by a processor, the steps of the authenticity identification method described in the foregoing method embodiments are executed.
  • the storage medium may be a volatile or non-volatile computer-readable storage medium.
  • the computer program product of the authenticity identification method provided by the embodiments of the present disclosure includes a computer-readable storage medium storing program codes, and the instructions included in the program code can be used to execute the authenticity identification method described in the above method embodiments
  • the computer program product can be specifically realized by means of hardware, software or a combination thereof.
  • the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (Software Development Kit, SDK) etc. wait.
  • the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit.
  • the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium executable by a processor.
  • the technical solution of the present disclosure is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present disclosure.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disc and other media that can store program codes. .

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

本公开提供了一种真伪识别方法、装置、电子设备以及存储介质，应用于目标深度神经网络。具体地：获取包括待识别对象的待识别图像；提取所述待识别图像对应于多个预设维度的图像特征；其中，图像特征对应的特征点的数量与对应的预设维度的值负相关；基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。

Description

真伪识别方法、装置、电子设备以及存储介质
本公开要求于2021年10月29日提交中国专利局、申请号为202111275648.6、发明名称为“真伪识别方法、装置、电子设备以及存储介质”，其全部内容通过引用结合在本公开中。
技术领域
本公开涉及图像处理技术领域，具体涉及一种真伪识别方法、装置、电子设备以及存储介质。
背景技术
近年来，随着数码相机、通信技术和移动设备的快速发展，图片和视频已经成为人们生活中的一种娱乐和交流方式。而与此同时，人脸造假技术充斥着网络，由于其极低的门槛和广泛的可及性，这项技术被恶意地应用于编造假新闻、进行人物抹黑等。人脸造假技术无限制的使用不仅危害隐私、安全等，而且还造成了社会上严重的信任危机。
因此，实现准确的人脸造假检测迫在眉睫，目前的人脸伪造识别准确度不高，并且随着伪造技术针对人脸伪造检测所做的改进，严重降低了当前的伪造检测精度。
发明内容
本公开实施例至少提供一种真伪识别方法、装置、电子设备以及存储介质。
第一方面，本公开实施例提供了一种真伪识别方法，应用于目标深度神经网络，包括：
获取包括待识别对象的待识别图像；
提取所述待识别图像对应于多个预设维度的图像特征；其中，图像特征对应的特征点的数量与对应的预设维度的值负相关；
基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。
该方面，利用目标深度神经网络能够较为准确地确定描述待识别图像的多个预设维度的图像特征，之后利用确定的图像特征能够较为准确地识别出待识别对象是否为伪造对象，即得到较为准确的真伪结果信息，以及，能够得到较为准确的伪造区域。进一步地，利用同一个目标神经网络同时对待识别对象进行真伪鉴别和确定图像中的伪造区域，这两个检测任务能够相互促进以及共用一些特定的图像特征，不仅能够提高检测效率，并且能够增强对图像特征的提取能力，提高两个检测任务的检测精度。
在一种可能的实施方式中，所述基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项，包括：
基于所述多个预设维度中最高预设维度对应的第一图像特征，得到所述待识别对象的真伪结果信息；
和/或，基于所述多个预设维度中最低预设维度对应的第二图像特征，确定所述待识别图像中的伪造区域。
该实施方式，最高预设维度的第一图像特征中特征点的数量较少，一个特征点对应的 特征数据维度较高，因此有效去除了待识别图像中的冗余信息，同时保留和增加了对真伪识别的有效信息，因此能够有效提高确定的真伪结果信息的准确性；最低预设维度对应的第二图像特征中特征点的数量多，因此能够较为准确地确定对应的待识别图像中每个像素点对应的真伪信息，继而能够较为准确地确定待识别图像中的伪造区域，另外，由于最低预设维度对应的第二图像特征中每个特征点对应的特征数据维度较低，因此能够提高确定伪造区域的速度。
在一种可能的实施方式中，所述图像特征包括多个预设维度中每个预设维度对应的第二图像特征，和多个预设维度中每个预设维度对应的第一图像特征；
所述提取所述待识别图像对应于多个预设维度的图像特征，包括：
提取所述待识别图像对应于多个预设维度中每个预设维度的第一图像特征；
基于各个预设维度对应的第一图像特征，确定所述待识别图像对应于多个预设维度中每个预设维度的第二图像特征。
该实施方式，对各个预设维度的第一图像特征进行降维和融合等操作能够较为准确地确定待识别图像对应于多个预设维度中每个预设维度的第二图像特征。
在一种可能的实施方式中，所述基于各个预设维度对应的第一图像特征，确定所述待识别图像对应于多个预设维度中每个预设维度的第二图像特征，包括：
将每两个相邻的预设维度作为一个组，得到多个第一维度组；所述第一维度组包括第一预设维度和第二预设维度，所述第一预设维度高于所述第二预设维度；
按照维度从高到低的顺序，分别对每个非最高维度的第一维度组执行如下操作，直到确定所述最低预设维度对应的第二图像特征：
按照所述第一维度组中的第二预设维度，对所述第一维度组中的第一预设维度对应的第二图像特征进行第一特征处理操作，得到与所述第二预设维度相匹配的第三图像特征；其中，所述第三图像特征对应的特征图与所述第二预设维度的第一图像特征对应的特征图具有相同的图像分辨率；
基于得到的所述第三图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
该实施方式，将相邻两个预设维度中较高的预设维度对应的第一图像特征进行处理，处理后的第三图像特征与较低的预设维度对应的第一图像特征具有相同的数据维度，并且对应的特征图具有相同的图像分辨率；后续基于数据维度和图像分辨率相同的第三图像特征和第一图像特征进行特征融合，能够提高融合精度，得到较为准确地第二图像特征。
在一种可能的实施方式中，所述基于得到的所述第三图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征，包括：
将得到的所述第三图像特征与所述第一维度组中的第二预设维度对应的第一图像特征进行拼接处理，得到拼接图像特征；
基于所述拼接图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
该实施方式，上述第三图像特征和第一图像特征具体有相同的数据维度以及相同的图像分辨率，因此可以准确地将两者进行拼接；之后，基于拼接后的拼接图像特征进行特征 提取和处理等操作，能够得到较为准确的处理结果，即上述第二图像特征。
在一种可能的实施方式中，所述提取所述待识别图像对应于多个预设维度中每个预设维度的第一图像特征，包括：
提取所述待识别图像对应于最低预设维度的第一图像特征；
将每两个相邻的预设维度作为一个组，得到多个第二维度组；所述第二维度组包括第三预设维度和第四预设维度，所述第三预设维度低于所述第四预设维度；
按照维度从低到高的顺序，分别对每个第二维度组执行如下操作，直到确定除最低预设维度以外的每个预设维度的第一图像特征：
按照所述第二维度组中的第四预设维度，对所述第二维度组中的第三预设维度对应的第一图像特征进行第二特征处理操作，得到与所述第四预设维度相匹配的第四图像特征；
基于得到的所述第四图像特征，确定所述第二维度组中的第四预设维度对应的第一图像特征。
该实施方式，按照第四预设维度对第一图像特征进行处理，能够较为准确地确定与第四预设维度相匹配的第四图像特征；后续再对该第四图像特征进行继续处理，得到的第一图像特征与第四预设维度相匹配。按照本实施方式依次对每个第二维度组中较低预设维度的第一图像特征进行处理，能够较为准确地确定每个预设维度对应的第一图像特征。
在一种可能的实施方式中，所述基于所述多个预设维度中最高预设维度对应的第一图像特征，得到所述待识别对象的真伪结果信息，包括：
基于最高预设维度对应的第一图像特征，确定所述待识别对象为真实对象的第一分值和所述待识别对象为伪造对象的第二分值；
基于所述第一分值和第二分值，确定所述待识别对象为真实对象的第一预测概率和所述待识别对象为伪造对象的第二预测概率；
基于所述第一预测概率和所述第二预测概率，确定所述待识别对象的真伪结果信息。
该实施方式，最高预设维度对应的第一图像特征中每个特征点对应的特征数据的维度较多，特征点的数量较少，这样的图像特征能够较为准确地描述对真伪识别有效的图像特征，因此利用第一图像特征能够较为准确地确定上述第一预测概率和第二预测概率，继而，基于该第一预测概率和第二预测概率，能够得到较为准确的真伪结果信息。
在一种可能的实施方式中，所述基于所述多个预设维度中最低预设维度对应的第二图像特征，确定所述待识别图像中的伪造区域，包括：
基于最低预设维度对应的第二图像特征，确定所述待识别图像中每个像素点为伪造像素点的概率信息；
基于确定的所述概率信息和预设概率阈值，确定所述待识别图像中每个像素点为伪造像素点的伪造结果信息；
基于每个像素点对应的所述伪造结果信息，确定所述待识别图像中的伪造区域。
该实施方式，最低预设维度对应的第二图像特征是各级预设维度对应的第一图像特征经过降维、增加对应的特征点以及拼接等操作得到的，因此，第二图像特征的每个特征点对应的特征数据能够较为准确的表征对应的特征点是否为伪造特征点；又由于该第二图像特征对应的特征图与待识别图像的分辨率相同，因此，利用该第二图像特征能够较为准确 地确定待识别图像中每个像素点为伪造像素点的伪造结果信息，继而能够确定较为精确的伪造区域。
在一种可能的实施方式中，所述获取包括待识别对象的待识别图像，包括：
获取原始图像；
对所述原始图像进行识别，确定所述待识别对象的检测框和所述待识别对象对应的多个关键点；
基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
该实施方式，检测框和关键点均能够分别来确定待识别对象在原始图像中所占的图像区域，将两者结合来确定上述图像区域，能够起到相互校准的作用，因此能够得到更加精准的图像区域，即能够得到更加精准的待识别图像。
在一种可能的实施方式中，所述基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像，包括：
基于所述多个关键点和所述检测框，确定所述待识别对象在所述原始图像中的初始区域信息；
按照预设比例信息，对所述初始区域信息对应的区域进行扩展，得到所述待识别对象在所述原始图像中的目标区域信息；
按照所述目标区域信息，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
该实施方式，通过对初始区域信息对应的区域进行扩展，能够保证目标区域信息对应的区域包括完整的待识别对象，有利于提高待识别对象对应的真伪识别精度。
在一种可能的实施方式中，所述基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像，包括：
确定所述待识别对象在所述原始图像中区域面积信息；
在所述区域面积信息对应的区域面积大于预设面积的情况下，对基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
该实施方式，提取在原始图像所占的面积较大的待识别对象对应的图像区域，能够保证得到的待识别图像的分辨率较大，有利于提高真伪识别精度。
在一种可能的实施方式中，在确定所述伪造区域之后，上述方法还包括：
基于所述伪造区域和所述待识别图像，生成热力图；其中，所述热力图中对应于所述伪造区域的像素点的热力值高于预设值。
该实施方式，利用热力图实现了对待识别图像对应的伪造区域的可视化，提高了伪造区域的直观性。
在一种可能的实施方式中，上述真伪识别方法还包括训练所述目标深度神经网络的步骤：
获取多张样本图像；
将所述样本图像输入待训练的目标深度神经网络，经过所述目标神经网络对所述多张 样本图像进行处理，得到每张样本图像中样本对象为真实对象的第一预测分值、所述样本对象为伪造对象的第二预测分值，和每张样本图像中每个像素点为伪造像素点的预测概率信息；
基于每张样本图像对应的第一预测分值、第二预测分值、预测概率信息和标准概率信息，生成网络损失信息；
利用所述网络损失信息对待训练的目标深度神经网络进行训练，直到满足预设训练条件，得到训练好的目标深度神经网络。
该实施方式，基于每个像素点对应的预测概率信息，能够直接得到伪造区域的预测结果，因此上述预测概率信息可以用于表征伪造区域的预测结果；通过第一预测分值和第二预测分值能够确定样本对象的真伪鉴别结果。因此，基于第一预测分值和第二预测分值(对应于真伪鉴别的检测任务)，以及，预测概率信息和标准概率信息(对应于伪造区域的检测任务)，这两种检测任务的预测值来建立训练目标神经网络的网络损失信息，能够通过两种检测任务的相互促进作用，有效提高训练完成的目标神经网络的检测精度。
在一种可能的实施方式中，所述基于每张样本图像对应的第一预测分值、第二预测分值、预测概率信息和标准概率信息，生成网络损失信息，包括：
基于每张样本图像对应的第一预测分值、第二预测分值，生成第一损失信息；
基于每张样本图像对应的预测概率信息和标准概率信息，生成第二损失信息；
基于所述第一损失信息和所述第二损失信息，生成所述网络损失信息。
该实施方式，利用对样本图像的真伪鉴别信息，即上述第一预测分值和第二预测分值，能够较为准确地确定上述第一损失信息；利用伪造区域对应的预测信息(即上述预测概率信息)和标准结果(标准概率信息)，能够较为准确地确定上述第二损失信息；之后，基于该第一损失信息和第二损失信息能够生成表征两种检测任务的损失的网络损失信息。
第二方面，本公开提供了一种真伪识别装置，包括：
图像获取模块，用于获取包括待识别对象的待识别图像；
特征提取模块，用于提取所述待识别图像对应于多个预设维度的第一图像特征；其中，第一图像特征对应的特征点的数量与对应的预设维度的值负相关；
检测模块，用于基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。
第三方面，本公开实施例还提供一种电子设备，包括：处理器、存储器和总线，所述存储器存储有所述处理器可执行的机器可读指令，当电子设备运行时，所述处理器与所述存储器之间通过总线通信，所述机器可读指令被所述处理器执行时执行上述第一方面，或第一方面中任一种可能的实施方式中的步骤。
第四方面，本公开实施例还提供一种计算机可读存储介质，该计算机可读存储介质上存储有计算机程序，该计算机程序被处理器运行时执行上述第一方面，或第一方面中任一种可能的实施方式中的步骤。
关于上述真伪识别方法装置、电子设备、及计算机可读存储介质的效果描述参见上述真伪识别方法的说明，这里不再赘述。
为使本公开的上述目的、特征和优点能更明显易懂，下文特举较佳实施例，并配合所 附附图，作详细说明如下。
附图说明
为了更清楚地说明本公开实施例的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，此处的附图被并入说明书中并构成本说明书中的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。应当理解，以下附图仅示出了本公开的某些实施例，因此不应被看作是对范围的限定，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他相关的附图。
图1示出了本公开实施例所提供的一种真伪识别方法的流程图；
图2示出了本公开实施例所提供的另一种真伪识别方法的流程图；
图3示出了本公开实施例所提供的网络训练方法的流程图；
图4示出了本公开实施例所提供的一种真伪识别装置的示意图；
图5示出了本公开实施例所提供的一种电子设备的示意图。
具体实施方式
为使本公开实施例的目的、技术方案和优点更加清楚，下面将结合本公开实施例中附图，对本公开实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本公开一部分实施例，而不是全部的实施例。通常在此处附图中描述和示出的本公开实施例的组件可以以各种不同的配置来布置和设计。因此，以下对在附图中提供的本公开的实施例的详细描述并非旨在限制要求保护的本公开的范围，而是仅仅表示本公开的选定实施例。基于本公开的实施例，本领域技术人员在没有做出创造性劳动的前提下所获得的所有其他实施例，都属于本公开保护的范围。
经研究发现，现有技术的人脸伪造识别中存在的检测精度低的缺陷，本公开提供了一种真伪识别方法、装置、电子设备以及存储介质。本公开利用目标深度神经网络能够较为准确地确定描述待识别图像的多个预设维度的图像特征，之后利用确定的图像特征能够较为准确地识别出待识别对象是否为伪造对象，即得到较为准确的真伪结果信息，以及，能够得到较为准确的伪造区域。进一步地，利用同一个目标神经网络同时对待识别对象进行真伪鉴别和确定图像中的伪造区域，这两个检测任务能够相互促进以及共用一些特定的图像特征，不仅能够提高检测效率，并且能够增强对图像特征的提取能力，提取到图像中对应于真伪鉴别的关键信息，提高两个检测任务的检测精度。
下面以执行主体为具有计算能够的设备为例对本公开实施例提供的真伪识别方法加以说明。
如图1所示，本公开提供的真伪识别方法应用于目标深度神经网络，可以包括如下步骤：
S110、获取包括待识别对象的待识别图像。
上述待识别对象可以是需要进行真伪鉴别的对象，例如，可以是人脸，待识别对象可以根据具体的应用场景来确定，本公开对此并不进行限定。
上述待识别图像可以是上述具有计算能够的设备拍摄的，也可以是其他拍摄设备拍摄 传送给上述计算能够的设备的，本公开对此并不进行限定。
上述待识别图像可以是拍摄的原始图像，也可以是从拍摄的原始图像中截取出来的待识别对象对应的子图像，本公开对此并不进行限定。
上述待识别图像可以是从视频片段中筛选出来的，也可以是独立存在的图像，本公开对此并不进行限定。
S120、提取所述待识别图像对应于多个预设维度的图像特征；其中，图像特征对应的特征点的数量与对应的预设维度的值负相关。
示例性地，上述图像特征可以包括多个预设维度中每个预设维度对应的第二图像特征，和多个预设维度中每个预设维度对应的第一图像特征。
示例性地，可以利用如下步骤提取上述图像特征：首先提取所述待识别图像对应于多个预设维度中每个预设维度的第一图像特征，之后，基于各个预设维度对应的第一图像特征，确定所述待识别图像对应于多个预设维度中每个预设维度的第二图像特征。
不同的预设维度对应的第一图像特征的特征点数量不同，特征点对应的特征数据的数据维度也不同。示例性地，预设维度越高，对应的第一图像特征的特征点的数量越少，一个特征点对应的特征数据的数据维度越高，对应的第一图像特征的特征图越小。
预设维度是预先根据具体的应用场景设定的。
示例性地，首先对待识别图像进行特征提取，得到与最低的预设维度对应原始图像特征，之后利用卷积层等对提取得到图像特征进行处理，得到与最低的预设维度对应第一图像特征。
之后，按照预设维度从低到高的顺序，根据较低的预设维度对应的第一图像特征，得到相邻的、较高预设维度对应的第一图像特征。
上述第二图像特征对应的特征图与上述待识别图像具有相同的尺寸和分辨率。
示例性地，按照预设维度从高到低的顺序，通过对第一图像特征进行降低特征数据的数据维度、拼接等操作，最终确定最低预设维度对应的第二图像特征。该第二图像特征的每个特征点对应的特征数据能够较为准确的表征对应的特征点是否为伪造特征点。
S130、基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。
示例性地，可以基于所述多个预设维度中最高预设维度对应的第一图像特征，得到所述待识别对象的真伪结果信息；和/或，基于所述多个预设维度中最低预设维度对应的第二图像特征，确定所述待识别图像中的伪造区域。
最高预设维度对应的第一图像特征中每个特征点对应的特征数据的维度较多，特征点的数量较少，这样的图像特征能够较为准确地描述对真伪识别有效的图像特征，因此利用第一图像特征能够较为准确地确定待识别对象的真伪结果信息。
示例性地，利用卷积层对上述最高预设维度对应的第一图像特征进行处理，得到上述真伪结果信息。
最低预设维度对应的第二图像特征是各级预设维度对应的第一图像特征经过降维、增加对应的特征点以及拼接等操作得到的，因此，第二图像特征的每个特征点对应的特征数据能够较为准确的表征对应的特征点是否为伪造特征点；又由于该第二图像特征对应的特 征图与待识别图像的分辨率相同，因此，利用该第二图像特征能够较为准确地确定待识别图像中每个像素点为伪造像素点的伪造结果信息，继而能够确定较为精确的伪造区域。
示例性地，利用卷积层对上述最低预设维度对应的第二图像特征进行处理，得到上述伪造区域。
在一些实施例中，可以利用如下步骤确定各个预设维度对应的第二图像特征：
首先，对预设维度进行分组：将每两个相邻的预设维度作为一个组，得到多个第一维度组；所述第一维度组包括第一预设维度和第二预设维度，所述第一预设维度高于所述第二预设维度。其中，某一第一维度组的维度可以是该第一维度组中的第一预设维度。
之后，按照维度从高到低的顺序，分别对每个第一维度组执行如下操作，直到确定所述最低预设维度对应的第二图像特征：
针对除最高维度的第一维度组以外的其他第一维度组，按照所述第一维度组中的第二预设维度，对该第一维度组中的第一预设维度对应的第二图像特征进行特征处理操作，得到与所述第二预设维度相匹配的第三图像特征；其中，所述第三图像特征对应的特征图与所述第二预设维度对应的第一图像特征对应的特征图具有相同的图像分辨率和尺寸。基于得到的所述第三图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
针对具有最高维度的第一维度组，对该第一维度组中的第一预设维度对应的第一图像特征进行特征处理操作，得到与所述第二预设维度相匹配的第五图像特征；其中，所述第五图像特征对应的特征图与所述第二预设维度对应的第一图像特征对应的特征图具有相同的图像分辨率和尺寸。基于得到的所述第五图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
示例性地，上述特征处理操作可以是转置卷积操作，通过该操作可以实现降低特征点对应的特征数据的数据维度，同时增加特征点的数量，即提升对应的特征图的分辨率。
如图2所示，对于最高维度的第一维度组C1，将该第一维度组中相邻两个预设维度中较高的预设维度对应的第一图像特征291进行特征处理操作，处理后的第五图像特征与较低的预设维度对应的第一图像特征具有相同的数据维度，并且对应的特征图具有相同的图像分辨率；之后，将该特征处理操作后的第三图像特征与该第一维度组中的第二预设维度对应的第一图像特征进行拼接处理，得到拼接图像特征，如图2所示拼接图像特征21。之后，基于拼接图像特征21，确定该第一维度组中的第二预设维度对应的第二图像特征23，例如可以对拼接图像特征21进行至少一个卷积处理，得到第二图像特征23。另外，如图2示出了拼接图像特征21、第二图像特征23等。
如图2所示，对于其他维度的第一维度组，例如第一维度组C2，将该第一维度组中相邻两个预设维度中较高的预设维度对应的第二图像特征23进行特征处理操作，处理后的第三图像特征与较低的预设维度对应的第一图像特征具有相同的数据维度，并且对应的特征图具有相同的图像分辨率；之后，将该第三图像特征与该第一维度组中的第二预设维度对应的第一图像特征进行拼接处理，得到拼接图像特征，如图2所示拼接图像特征22。之后，基于拼接图像特征22，确定该第一维度组中的第二预设维度对应的第二图像特征24，例如可以对拼接图像特征22进行至少一个卷积处理，得到第二图像特征24。
上述同一第一维度组对应的第三图像特征和第二预设维度对应的第一图像特征具体有相同的数据维度以及相同的图像分辨率，因此可以准确地将两者进行拼接；之后，基于拼接后的拼接图像特征进行特征提取和处理等操作，能够得到较为准确的处理结果，即上述第二图像特征。
根据上述实施例可知，按照维度从高到低的顺序依次对每个第一维度组中的第一图像特征和第二图像特征进行处理，可以得到最低预设维度对应的第二图像特征。在得到最低预设维度对应的第二图像特征之后，可以利用如下步骤确定待识别图像中的伪造区域：
首先，基于最低预设维度对应的第二图像特征，确定所述待识别图像中每个像素点为伪造像素点的概率信息。
如图2示出了最低预设维度的第二图像特征25。
最低预设维度的第二图像特征对应的特征图与待识别图像的分辨率和尺寸均相同，因此，该第二图像特征对应的特征图中的每个像素点与待识别图像中的每个像素点相对应，那么，利用该第二图像特征能够较为准确地确定待识别图像中每个像素点为伪造像素点的伪造结果信息。
示例性地，可以利用全连接网络层、分类器等对最低预设维度对应的第二图像特征25进行处理，得到待识别图像中每个像素点为伪造像素点的概率信息。
在得到待识别图像中每个像素点对应的概率信息之后，基于确定的所述概率信息和预设概率阈值，确定所述待识别图像中每个像素点为伪造像素点的伪造结果信息。基于每个像素点对应的所述伪造结果信息，确定所述待识别图像中的伪造区域。
上述预设概率阈值是根据具体的应用场景灵活设定的。
示例性地，在上述概率信息对应的概率值大于或等于上述预设概率阈值时，确定上述概率信息对应的像素点为伪造像素点；在上述概率信息对应的概率值小于上述预设概率阈值时，确定上述概率信息对应的像素点为未经过篡改的像素点。
在确定了各个像素点是否为伪造像素点之后，确定为伪造像素点的像素点可以形成至少一个伪造区域。
示例性地，确定了各个像素点是否为伪造像素点之后，创建一个与待识别图像相同尺寸的掩膜MASK图M pred。之后，根据下面的公式填充MASK图M pred：
Figure PCTCN2022096019-appb-000001
式中，(i，j)表示对应像素点的行、列标识符，λ表示上述预设概率阈值，I pred表示上述概率信息对应的概率值。
在确定所述伪造区域之后，可以基于伪造区域和待识别图像，生成热力图；其中，所述热力图中对应于所述伪造区域的像素点的热力值高于预设值。
示例性地，根据待识别图像的分辨率和尺寸，确定热力图的分辨率和尺寸，例如，可以将热力图设置为与待识别图像具有相同的尺寸和分辨率。之后，可以将热力图中对应于所述伪造区域的像素点的热力值高于预设值，并且热力图中的该部分像素点的热力值相等；也可以根据伪造区域内每个像素点对应的上述概率信息，设置热力图中对应的像素点的热力值，具体地，随着概率信息对应的概率值增高，对应的像素点的热力值也增高。
热力图中，对应于伪造区域以外的像素点的热力值可以设置为相等，也可以根据上述概率信息对应的概率值进行设定，本公开并不进行限定。
利用热力图实现了对待识别图像对应的伪造区域的可视化，提高了伪造区域的直观性和解释性。
在一些实施例中，可以利用如下步骤提取所述待识别图像对应于多个预设维度中每个预设维度的第一图像特征：
首先提取最低预设维度对应的第一图像特征，示例性地，可以将待识别图像输入目标深度神经网络，之后经过至少一次列深度可分离卷积操作，得到上述最低预设维度的第一图像特征26。
在得到最低预设维度的第一图像特征之后，可以将每两个相邻的预设维度作为一个组，得到多个第二维度组；所述第二维度组包括第三预设维度和第四预设维度，所述第三预设维度低于所述第四预设维度。其中，第二维度组的维度可以为该第二维度组中的第三预设维度。
确定好上述第二维度组之后，按照维度从低到高的顺序，依次对每个第二维度组执行如下操作，直到确定除最低预设维度以外的每个预设维度的第一图像特征：
按照所述第二维度组中的第四预设维度，对所述第二维度组中的第三预设维度对应的第一图像特征进行第二特征处理操作，得到与所述第四预设维度相匹配的第四图像特征；基于得到的所述第四图像特征，确定所述第二维度组中的第四预设维度对应的第一图像特征。
示例性地，如图2所示，第二维度组中较低的预设维度对应的第一图像特征26，进行特征处理操作，得到与所述较高的预设维度相匹配的第四图像特征29，之后，基于得到的所述第四图像特征29，经过卷积等操作，确定较高的预设维度对应的第一图像特征27。
示例性地，上述第二特征处理操作可以是可分离卷积操作，具体用于升高第一图像特征中特征点的特征数据的数据维度，并减少特征点的数量。经过第二特征处理操作之后得到的第四图像特征，与该第二维度组中较高的第四预设维度相匹配。
针对每个第二维度组，在得到该第二维度组对应的第四图像特征之后，可以经过至少一次卷积操作，得到该第二维度组中较高的预设维度对应的第一图像特征，该第一图像特征与第四图像特征具有相同的预设维度，并且对应的特征图具有相同的图像分辨率。
按照每个第二维度组中第四预设维度对对应第二维度组中的第一图像特征进行处理，能够较为准确地确定与第四预设维度相匹配的第四图像特征；后续再对该第四图像特征进行继续处理，得到的第一图像特征与第四预设维度相匹配。按照本实施方式依次对每个第二维度组中较低预设维度的第一图像特征进行处理，能够较为准确地确定每个预设维度对应的第一图像特征。
如图2示出第二维度组中的第一图像特征27、第一图像特征28等，以及第二维度组中的第四图像特征29等。
在得到最高预设维度对应的第一图像特征之后，可以利用如下步骤确定所述待识别对象的真伪结果信息：
首先确定两个分值，具体地，基于最高预设维度对应的第一图像特征，确定所述待识 别对象为真实对象的第一分值和所述待识别对象为伪造对象的第二分值。
示例性地，可以经过至少一次卷积操作对第一图像特征进行处理，得到待识别对象为真实对象的第一分值和待识别对象为伪造对象的第二分值。
得到上述两个分值之后确定两个预测概率，具体地，基于所述第一分值和第二分值，确定所述待识别对象为真实对象的第一预测概率和所述待识别对象为伪造对象的第二预测概率。
示例性地，可利用如下公式确定上述两个预测概率：
Figure PCTCN2022096019-appb-000002
式中，i、class均表示待识别对象为真实对象或伪造对象的标识，i为0表示待识别对象为真实对象，i为1表示待识别对象为伪造对象，p表示预测概率，class为0时，p class表示待识别对象为真实对象的第一预测概率，class为1时，p class表示待识别对象为伪造对象的第二预测概率，x表示分值，i为0时，x[i]表示待识别对象为真实对象的第一分值，i为1时，x[i]表示待识别对象为伪造对象的第二分值。
示例性地，可以利用分类器确定上述两个预测概率。
在得到上述两个预测概率之后，进行真伪结果预测，具体地，基于所述第一预测概率和所述第二预测概率，确定所述待识别对象的真伪结果信息。
示例性地，可以将第一预测概率和第二预测概率进行比较，并将较大的预测概率对应的鉴别结果作为上述真伪结果信息，例如，若第一预测概率大于第二预测概率，则真伪结果信息为待识别对象为真实对象，若第一预测概率小于或等于第二预测概率，则真伪结果信息为待识别对象为伪造对象。
具体地，可以利用如下公式实现：
c＝argmax(p 0,p 1)  (3)
式中，c为用于表征真伪结果信息的参数。
上述实施例中，最高预设维度对应的第一图像特征中每个特征点对应的特征数据的维度较多，特征点的数量较少，这样的图像特征能够较为准确地描述对真伪识别有效的图像特征，因此利用第一图像特征能够较为准确地确定上述第一预测概率和第二预测概率，继而，基于该第一预测概率和第二预测概率，能够得到较为准确的真伪结果信息。
上述实施例中的待识别图像既可以是拍摄装置直接拍摄得到的，也可以是从拍摄的原始图像中截取出来的子图像。示例性地，可以利用如下步骤截取待识别对象对应的待识别图像：
首先，对获取的原始图像进行识别，确定所述待识别对象的检测框和所述待识别对象对应的多个关键点；之后，基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
上述原始图像可以是从一视频片段中截取得到的，示例性地，可以从一视频片段中等间隔的进行抽帧处理，得到多帧原始图像，之后对每帧原始图像实施本公开的方法，可以实现对每帧图像的真伪鉴别。
检测框和关键点均能够分别来确定待识别对象在原始图像中所占的图像区域，将两者 结合来确定上述图像区域，能够起到相互校准的作用，因此能够得到更加精准的图像区域，即能够得到更加精准的待识别图像。
示例性地，首先统计位于检测框内的关键点的数量；之后，基于统计得到的数量，确定位于检测框内的关键点在所有关键点中的占比，在所述占比大于预设占比阈值的情况下，可以基于关键点的位置和检测框的位置，确定上述图像区域。确定的图像区域可以只是检测框对应的区域，可以是包括检测框对应的区域和所有关键点的区域。
在确定的上述占比小于或等于预设占比阈值的情况下，说明对象识别的误差较大，此时可以重新进行对象识别，重新确定待识别对象的检测框和关键点。
在一些实施例中，上述基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像，具体可以利用如下步骤实现：
首先基于所述多个关键点和所述检测框，确定所述待识别对象在所述原始图像中的初始区域信息；之后，按照预设比例信息，对所述初始区域信息对应的区域进行扩展，得到所述待识别对象在所述原始图像中的目标区域信息；按照所述目标区域信息，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
上述初始区域信息对应的区域可以是包括检测框和预设数量个关键点的区域。
通过对初始区域信息对应的区域进行扩展，能够保证目标区域信息对应的区域包括完整的待识别对象以及对象周围的少部分环境，有利于提高待识别对象对应的真伪识别精度。
示例性地，在对待识别图像进行识别时，不仅能够得到检测框和关键点，还能够确定待识别对象在原始图像中所占区域的面积(下述称为区域面积信息)以及待识别对象的姿态信息，这些信息可以保存在json文件，供后续处理。在需要进行真伪鉴别时，从json文件中提取所需的信息即可。
示例性地，在提取待识别图像时，首先可以从json文件中获取区域面积信息，在所述区域面积信息对应的区域面积大于预设面积的情况下，对基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
得到的待识别图像可以保存为png格式的图片。
提取在原始图像所占的面积较大的待识别对象对应的图像区域，能够保证得到的待识别图像的分辨率较大，有利于提高真伪识别精度。
本公开还提供了一种目标深度神经网络的训练方法，如图3所示，可以包括如下步骤：
S310、获取多张样本图像。
上述样本图像为包括样本对象的图像，例如包括人脸的样本图像。这里的样本图像既可以是拍摄设备拍摄的原始图像，也可以是从原始图像中截取的包括待识别对象的子图像。
S320、将所述样本图像输入待训练的目标深度神经网络，经过所述目标神经网络对所述多张样本图像进行处理，得到每张样本图像中样本对象为真实对象的第一预测分值、所述样本对象为伪造对象的第二预测分值，和每张样本图像中每个像素点为伪造像素点的预测概率信息。
S330、基于每张样本图像对应的第一预测分值、第二预测分值、预测概率信息和标准概率信息，生成网络损失信息。
上述网络损失信息包括对样本对象进行真伪检测对应的第一损失信息和确定损失区域 对应的第二损失信息。
根据上述第一预测分值和第二预测分值，能够确定样本对象为真实对象的第一样本概率和样本对象为伪造对象的第二样本概率。这里第一样本概率和第二样本概率的计算方法与上述第一预测概率和第二预测概率的计算方法相同，这里不再赘述。
示例性地，基于每张样本图像对应的第一样本概率、第二样本概率，生成第一损失信息，可以利用如下公式实现：
Figure PCTCN2022096019-appb-000003
式中，L c表示第一损失信息，i表示样本对象为真实对象或伪造对象的标识，i为0表示样本对象为真实对象，i为1表示样本为伪造对象，p表示样本概率，p 0表示样本对象为真实对象的第一样本概率，p 1表示样本对象为伪造对象的第二样本概率，q表示标准概率，q 0表示样本对象为真实对象的第一标准概率，q 1表示样本对象为伪造对象的第二标准概率。
示例性地，基于每张样本图像对应的预测概率信息和标准概率信息，生成第二损失信息，可以利用如下公式实现：
L region＝∑ i,j(M target(i,j)·log(I pred(i,j))+(1-M target(i,j))·log(1-I pred(i,j)))  (5)
式中，L region表示第二损失信息，(i，j)表示对应像素点的行、列标识符，I pred表示对应像素点对应的预测概率信息，Mtarget表示对应像素点对应的标准概率信息。
在得到上述第一损失信息和第二损失信息之后，可以基于所述第一损失信息和所述第二损失信息，生成所述网络损失信息。示例性地，可以利用如下公式实现：
L＝a×L c+b×L region  (6)
式中，L表示网络损失信息，a、b表示预设权重。
S340、利用所述网络损失信息对待训练的目标深度神经网络进行训练，直到满足预设训练条件，得到训练好的目标深度神经网络。
上述实施例中，基于每个像素点对应的预测概率信息，能够直接得到伪造区域的预测结果，因此上述预测概率信息可以用于表征伪造区域的预测结果；通过第一预测分值和第二预测分值能够确定样本对象的真伪鉴别结果。因此，基于第一预测分值和第二预测分值(对应于真伪鉴别的检测任务)，以及，预测概率信息和标准概率信息(对应于伪造区域的检测任务)，这两种检测任务的预测值来建立训练目标神经网络的网络损失信息，能够通过两种检测任务的相互促进作用，有效提高训练完成的目标神经网络的检测精度。
本领域技术人员可以理解，在具体实施方式的上述方法中，各步骤的撰写顺序并不意味着严格的执行顺序而对实施过程构成任何限定，各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
基于同一发明构思，本公开实施例中还提供了真伪识别方法对应的真伪识别装置，由于本公开实施例中的装置解决问题的原理与本公开实施例上述真伪识别方法相似，因此装置的实施可以参见方法的实施，重复之处不再赘述。
参照图4所示，为本公开实施例提供的一种真伪识别装置的架构示意图，所述装置包括：
图像获取模块410，用于获取包括待识别对象的待识别图像。
特征提取模块420，用于提取所述待识别图像对应于多个预设维度的图像特征；其中，图像特征对应的特征点的数量与对应的预设维度的值负相关。
检测模块430，用于基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。
在一些实施例中，所述检测模块430在所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项时，具体用于：
基于所述多个预设维度中最高预设维度对应的第一图像特征，得到所述待识别对象的真伪结果信息；
和/或，基于所述多个预设维度中最低预设维度对应的第二图像特征，确定所述待识别图像中的伪造区域。
在一些实施例中，所述图像特征包括多个预设维度中每个预设维度对应的第二图像特征，和多个预设维度中每个预设维度对应的第一图像特征；
所述特征提取模块420在提取所述待识别图像对应于多个预设维度的图像特征时，用于：
提取所述待识别图像对应于多个预设维度中每个预设维度的第一图像特征；
基于各个预设维度对应的第一图像特征，确定所述待识别图像对应于多个预设维度中每个预设维度的第二图像特征。
在一些实施例中，所述特征提取模块420在基于各个预设维度对应的第一图像特征，确定所述待识别图像对应于多个预设维度中每个预设维度的第二图像特征时，具体用于：
将每两个相邻的预设维度作为一个组，得到多个第一维度组；所述第一维度组包括第一预设维度和第二预设维度，所述第一预设维度高于所述第二预设维度；
按照维度从高到低的顺序，分别对每个非最高维度的第一维度组执行如下操作，直到确定所述最低预设维度对应的第二图像特征：
按照所述第一维度组中的第二预设维度，对所述第一维度组中的第一预设维度对应的第一图像特征进行第一特征处理操作，得到与所述第二预设维度相匹配的第三图像特征；其中，所述第三图像特征对应的特征图与所述第二预设维度的第一图像特征对应的特征图具有相同的图像分辨率；
基于得到的所述第三图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
在一些实施例中，所述特征提取模块420在基于得到的所述第三图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征时，用于：
将得到的所述第三图像特征与所述第一维度组中的第二预设维度对应的第一图像特征进行拼接处理，得到拼接图像特征；
基于所述拼接图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
在一些实施例中，所述特征提取模块420在提取所述待识别图像对应于多个预设维度中每个预设维度的第一图像特征时，用于：
提取所述待识别图像对应于最低预设维度的第一图像特征；
将每两个相邻的预设维度作为一个组，得到多个第二维度组；所述第二维度组包括第三预设维度和第四预设维度，所述第三预设维度低于所述第四预设维度；
按照维度从低到高的顺序，分别对每个第二维度组执行如下操作，直到确定除最低预设维度以外的每个预设维度的第一图像特征：
按照所述第二维度组中的第四预设维度，对所述第二维度组中的第三预设维度对应的第一图像特征进行第二特征处理操作，得到与所述第四预设维度相匹配的第四图像特征；
基于得到的所述第四图像特征，确定所述第二维度组中的第四预设维度对应的第一图像特征。
在一些实施例中，所述检测模块430在基于所述多个预设维度中最高预设维度对应的第一图像特征，得到所述待识别对象的真伪结果信息时，用于：
基于最高预设维度对应的第一图像特征，确定所述待识别对象为真实对象的第一分值和所述待识别对象为伪造对象的第二分值；
基于所述第一分值和第二分值，确定所述待识别对象为真实对象的第一预测概率和所述待识别对象为伪造对象的第二预测概率；
基于所述第一预测概率和所述第二预测概率，确定所述待识别对象的真伪结果信息。
在一些实施例中，所述检测模块430在基于所述多个预设维度中最低预设维度对应的第二图像特征，确定所述待识别图像中的伪造区域时，用于：
基于最低预设维度对应的第二图像特征，确定所述待识别图像中每个像素点为伪造像素点的概率信息；
基于确定的所述概率信息和预设概率阈值，确定所述待识别图像中每个像素点为伪造像素点的伪造结果信息；
基于每个像素点对应的所述伪造结果信息，确定所述待识别图像中的伪造区域。
在一些实施例中，所述图像获取模块410在获取包括待识别对象的待识别图像时，用于：
获取原始图像；
对所述原始图像进行识别，确定所述待识别对象的检测框和所述待识别对象对应的多个关键点；
基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
在一些实施例中，所述图像获取模块410在基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像时，用于：
基于所述多个关键点和所述检测框，确定所述待识别对象在所述原始图像中的初始区域信息；
按照预设比例信息，对所述初始区域信息对应的区域进行扩展，得到所述待识别对象在所述原始图像中的目标区域信息；
按照所述目标区域信息，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
在一些实施例中，所述图像获取模块410在基于所述检测框和所述多个关键点，从所 述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像时，用于：
确定所述待识别对象在所述原始图像中区域面积信息；
在所述区域面积信息对应的区域面积大于预设面积的情况下，对基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
在一些实施例中，在确定所述伪造区域之后，所述检测模块430还用于：
基于所述伪造区域和所述待识别图像，生成热力图；其中，所述热力图中对应于所述伪造区域的像素点的热力值高于预设值。
在一些实施例中，上述装置还包括对所述目标深度神经网络进行训练的训练模块440，所述训练模块440用于：
获取多张样本图像；
将所述样本图像输入待训练的目标深度神经网络，经过所述目标神经网络对所述多张样本图像进行处理，得到每张样本图像中样本对象为真实对象的第一预测分值、所述样本对象为伪造对象的第二预测分值，和每张样本图像中每个像素点为伪造像素点的预测概率信息；
基于每张样本图像对应的第一预测分值、第二预测分值、预测概率信息和标准概率信息，生成网络损失信息；
利用所述网络损失信息对待训练的目标深度神经网络进行训练，直到满足预设训练条件，得到训练好的目标深度神经网络。
在一些实施例中，所述训练模块440在基于每张样本图像对应的第一预测分值、第二预测分值、预测概率信息和标准概率信息，生成网络损失信息时，用于：
基于每张样本图像对应的第一预测分值、第二预测分值，生成第一损失信息；
基于每张样本图像对应的预测概率信息和标准概率信息，生成第二损失信息；
基于所述第一损失信息和所述第二损失信息，生成所述网络损失信息。
基于同一技术构思，本公开实施例还提供了一种电子设备。参照图5所示，为本公开实施例提供的电子设备500的结构示意图，包括处理器51、存储器52、和总线53。其中，存储器52用于存储执行指令，包括内存521和外部存储器522；这里的内存521也称内存储器，用于暂时存放处理器51中的运算数据，以及与硬盘等外部存储器522交换的数据，处理器51通过内存521与外部存储器522进行数据交换，当电子设备500运行时，处理器51与存储器52之间通过总线53通信，使得处理器51在执行以下指令：
获取包括待识别对象的待识别图像；提取所述待识别图像对应于多个预设维度的图像特征；其中，图像特征对应的特征点的数量与对应的预设维度的值负相关；基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。
本公开实施例还提供一种计算机可读存储介质，该计算机可读存储介质上存储有计算机程序，该计算机程序被处理器运行时执行上述方法实施例中所述的真伪识别方法的步骤。其中，该存储介质可以是易失性或非易失的计算机可读取存储介质。
本公开实施例所提供的真伪识别方法的计算机程序产品，包括存储了程序代码的计算机可读存储介质，所述程序代码包括的指令可用于执行上述方法实施例中所述的真伪识别 方法的步骤，具体可参见上述方法实施例，在此不再赘述。该计算机程序产品可以具体通过硬件、软件或其结合的方式实现。在一个可选实施例中，所述计算机程序产品具体体现为计算机存储介质，在另一个可选实施例中，计算机程序产品具体体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统和装置的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。在本公开所提供的几个实施例中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。以上所描述的装置实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，又例如，多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些通信接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本公开各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。
所述功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个处理器可执行的非易失的计算机可读取存储介质中。基于这样的理解，本公开的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本公开各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
最后应说明的是：以上所述实施例，仅为本公开的具体实施方式，用以说明本公开的技术方案，而非对其限制，本公开的保护范围并不局限于此，尽管参照前述实施例对本公开进行了详细的说明，本领域的普通技术人员应当理解：任何熟悉本技术领域的技术人员在本公开揭露的技术范围内，其依然可以对前述实施例所记载的技术方案进行修改或可轻易想到变化，或者对其中部分技术特征进行等同替换；而这些修改、变化或者替换，并不使相应技术方案的本质脱离本公开实施例技术方案的精神和范围，都应涵盖在本公开的保护范围之内。因此，本公开的保护范围应所述以权利要求的保护范围为准。

Claims (17)

  1. 一种真伪识别方法，其特征在于，应用于目标深度神经网络，包括：
    获取包括待识别对象的待识别图像；
    提取所述待识别图像对应于多个预设维度的图像特征；其中，图像特征对应的特征点的数量与对应的预设维度的值负相关；
    基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。
  2. 根据权利要求1所述的方法，其特征在于，所述基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项，包括：
    基于所述多个预设维度中最高预设维度对应的第一图像特征，得到所述待识别对象的真伪结果信息；
    和/或，基于所述多个预设维度中最低预设维度对应的第二图像特征，确定所述待识别图像中的伪造区域。
  3. 根据权利要求1或2所述的方法，其特征在于，所述图像特征包括多个预设维度中每个预设维度对应的第二图像特征，和多个预设维度中每个预设维度对应的第一图像特征；
    所述提取所述待识别图像对应于多个预设维度的图像特征，包括：
    提取所述待识别图像对应于多个预设维度中每个预设维度的第一图像特征；
    基于各个预设维度对应的第一图像特征，确定所述待识别图像对应于多个预设维度中每个预设维度的第二图像特征。
  4. 根据权利要求3所述的方法，其特征在于，所述基于各个预设维度对应的第一图像特征，确定所述待识别图像对应于多个预设维度中每个预设维度的第二图像特征，包括：
    将每两个相邻的预设维度作为一个组，得到多个第一维度组；所述第一维度组包括第一预设维度和第二预设维度，所述第一预设维度高于所述第二预设维度；
    按照维度从高到低的顺序，分别对非最高维度的第一维度组执行如下操作，直到确定所述最低预设维度对应的第二图像特征：
    按照所述第一维度组中的第二预设维度，对所述第一维度组中的第一预设维度对应的第二图像特征进行第一特征处理操作，得到与所述第二预设维度相匹配的第三图像特征；其中，所述第三图像特征对应的特征图与所述第二预设维度的第一图像特征对应的特征图具有相同的图像分辨率；
    基于得到的所述第三图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
  5. 根据权利要求4所述的方法，其特征在于，所述基于得到的所述第三图像特征和所述第一维度组中的第二预设维度对应的第一图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征，包括：
    将得到的所述第三图像特征与所述第一维度组中的第二预设维度对应的第一图像特征进行拼接处理，得到拼接图像特征；
    基于所述拼接图像特征，确定所述第一维度组中的第二预设维度对应的第二图像特征。
  6. 根据权利要求3至5任一项所述的方法，其特征在于，所述提取所述待识别图像对 应于多个预设维度中每个预设维度的第一图像特征，包括：
    提取所述待识别图像对应于最低预设维度的第一图像特征；
    将每两个相邻的预设维度作为一个组，得到多个第二维度组；所述第二维度组包括第三预设维度和第四预设维度，所述第三预设维度低于所述第四预设维度；
    按照维度从低到高的顺序，分别对每个第二维度组执行如下操作，直到确定除最低预设维度以外的每个预设维度的第一图像特征：
    按照所述第二维度组中的第四预设维度，对所述第二维度组中的第三预设维度对应的第一图像特征进行第二特征处理操作，得到与所述第四预设维度相匹配的第四图像特征；
    基于得到的所述第四图像特征，确定所述第二维度组中的第四预设维度对应的第一图像特征。
  7. 根据权利要求2所述的方法，其特征在于，所述基于所述多个预设维度中最高预设维度对应的第一图像特征，得到所述待识别对象的真伪结果信息，包括：
    基于最高预设维度对应的第一图像特征，确定所述待识别对象为真实对象的第一分值和所述待识别对象为伪造对象的第二分值；
    基于所述第一分值和第二分值，确定所述待识别对象为真实对象的第一预测概率和所述待识别对象为伪造对象的第二预测概率；
    基于所述第一预测概率和所述第二预测概率，确定所述待识别对象的真伪结果信息。
  8. 根据权利要求2或7所述的方法，其特征在于，所述基于所述多个预设维度中最低预设维度对应的第二图像特征，确定所述待识别图像中的伪造区域，包括：
    基于最低预设维度对应的第二图像特征，确定所述待识别图像中每个像素点为伪造像素点的概率信息；
    基于确定的所述概率信息和预设概率阈值，确定所述待识别图像中每个像素点为伪造像素点的伪造结果信息；
    基于每个像素点对应的所述伪造结果信息，确定所述待识别图像中的伪造区域。
  9. 根据权利要求1至8任一项所述的方法，其特征在于，所述获取包括待识别对象的待识别图像，包括：
    获取原始图像；
    对所述原始图像进行识别，确定所述待识别对象的检测框和所述待识别对象对应的多个关键点；
    基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
  10. 根据权利要求9所述的方法，其特征在于，所述基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像，包括：
    基于所述多个关键点和所述检测框，确定所述待识别对象在所述原始图像中的初始区域信息；
    按照预设比例信息，对所述初始区域信息对应的区域进行扩展，得到所述待识别对象在所述原始图像中的目标区域信息；
    按照所述目标区域信息，从所述原始图像中提取所述待识别对象对应的图像区域，得 到所述待识别图像。
  11. 根据权利要求9所述的方法，其特征在于，所述基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像，包括：
    确定所述待识别对象在所述原始图像中区域面积信息；
    在所述区域面积信息对应的区域面积大于预设面积的情况下，对基于所述检测框和所述多个关键点，从所述原始图像中提取所述待识别对象对应的图像区域，得到所述待识别图像。
  12. 根据权利要求1至11任一项所述的方法，其特征在于，在确定所述伪造区域之后，还包括：
    基于所述伪造区域和所述待识别图像，生成热力图；其中，所述热力图中对应于所述伪造区域的像素点的热力值高于预设值。
  13. 根据权利要求1至12任一项所述的方法，其特征在于，还包括训练所述目标深度神经网络的步骤：
    获取多张样本图像；
    将所述样本图像输入待训练的目标深度神经网络，经过所述目标神经网络对所述多张样本图像进行处理，得到每张样本图像中样本对象为真实对象的第一预测分值、所述样本对象为伪造对象的第二预测分值，和每张样本图像中每个像素点为伪造像素点的预测概率信息；
    基于每张样本图像对应的第一预测分值、第二预测分值、预测概率信息和标准概率信息，生成网络损失信息；
    利用所述网络损失信息对待训练的目标深度神经网络进行训练，直到满足预设训练条件，得到训练好的目标深度神经网络。
  14. 根据权利要求13所述的方法，其特征在于，所述基于每张样本图像对应的第一预测分值、第二预测分值、预测概率信息和标准概率信息，生成网络损失信息，包括：
    基于每张样本图像对应的第一预测分值、第二预测分值，生成第一损失信息；
    基于每张样本图像对应的预测概率信息和标准概率信息，生成第二损失信息；
    基于所述第一损失信息和所述第二损失信息，生成所述网络损失信息。
  15. 一种真伪识别装置，其特征在于，包括：
    图像获取模块，用于获取包括待识别对象的待识别图像；
    特征提取模块，用于提取所述待识别图像对应于多个预设维度的图像特征；其中，图像特征对应的特征点的数量与对应的预设维度的值负相关；
    检测模块，用于基于所述图像特征，确定所述待识别图像中的伪造区域和所述待识别对象的真伪结果信息中的至少一项。
  16. 一种电子设备，其特征在于，包括：处理器、存储器和总线，所述存储器存储有所述处理器可执行的机器可读指令，当电子设备运行时，所述处理器与所述存储器之间通过总线通信，所述机器可读指令被所述处理器执行时执行如权利要求1至14任一项所述的真伪识别方法的步骤。
  17. 一种计算机可读存储介质，其特征在于，该计算机可读存储介质上存储有计算机 程序，该计算机程序被处理器运行时执行如权利要求1至14任一项所述的真伪识别方法的步骤。
PCT/CN2022/096019 2021-10-29 2022-05-30 真伪识别方法、装置、电子设备以及存储介质 WO2023071180A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111275648.6A CN113920565A (zh) 2021-10-29 2021-10-29 真伪识别方法、装置、电子设备以及存储介质
CN202111275648.6 2021-10-29

Publications (1)

Publication Number Publication Date
WO2023071180A1 true WO2023071180A1 (zh) 2023-05-04

Family

ID=79244013

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/096019 WO2023071180A1 (zh) 2021-10-29 2022-05-30 真伪识别方法、装置、电子设备以及存储介质

Country Status (2)

Country Link
CN (1) CN113920565A (zh)
WO (1) WO2023071180A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113920565A (zh) * 2021-10-29 2022-01-11 上海商汤智能科技有限公司 真伪识别方法、装置、电子设备以及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130163829A1 (en) * 2011-12-21 2013-06-27 Electronics And Telecommunications Research Institute System for recognizing disguised face using gabor feature and svm classifier and method thereof
CN110555481A (zh) * 2019-09-06 2019-12-10 腾讯科技（深圳）有限公司 一种人像风格识别方法、装置和计算机可读存储介质
CN111310616A (zh) * 2020-02-03 2020-06-19 北京市商汤科技开发有限公司 图像处理方法及装置、电子设备和存储介质
CN112668462A (zh) * 2020-12-25 2021-04-16 平安科技（深圳）有限公司 车损检测模型训练、车损检测方法、装置、设备及介质
CN113920565A (zh) * 2021-10-29 2022-01-11 上海商汤智能科技有限公司 真伪识别方法、装置、电子设备以及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130163829A1 (en) * 2011-12-21 2013-06-27 Electronics And Telecommunications Research Institute System for recognizing disguised face using gabor feature and svm classifier and method thereof
CN110555481A (zh) * 2019-09-06 2019-12-10 腾讯科技（深圳）有限公司 一种人像风格识别方法、装置和计算机可读存储介质
CN111310616A (zh) * 2020-02-03 2020-06-19 北京市商汤科技开发有限公司 图像处理方法及装置、电子设备和存储介质
CN112668462A (zh) * 2020-12-25 2021-04-16 平安科技（深圳）有限公司 车损检测模型训练、车损检测方法、装置、设备及介质
CN113920565A (zh) * 2021-10-29 2022-01-11 上海商汤智能科技有限公司 真伪识别方法、装置、电子设备以及存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
WEN WU, ZHANG XU-HONG, WEN ZHI-YUN: "Text Feature Selection based on Improved CHI and PCA", COMPUTER ENGINEERING & SCIENCE, vol. 43, no. 9, 1 September 2021 (2021-09-01), XP093061965 *

Also Published As

Publication number Publication date
CN113920565A (zh) 2022-01-11

Similar Documents

Publication Publication Date Title
Zhou et al. Two-stream neural networks for tampered face detection
US11830230B2 (en) Living body detection method based on facial recognition, and electronic device and storage medium
WO2019218824A1 (zh) 一种移动轨迹获取方法及其设备、存储介质、终端
CN108805047B (zh) 一种活体检测方法、装置、电子设备和计算机可读介质
CN109858371B (zh) 人脸识别的方法及装置
Tariq et al. Gan is a friend or foe? a framework to detect various fake face images
US20200364478A1 (en) Method and apparatus for liveness detection, device, and storage medium
WO2022161286A1 (zh) 图像检测方法、模型训练方法、设备、介质及程序产品
WO2020258667A1 (zh) 图像识别方法及装置、非易失性可读存储介质、计算机设备
CN110853033B (zh) 基于帧间相似度的视频检测方法和装置
KR20200046181A (ko) 컨볼루션 뉴럴 네트워크를 통해 이미지 위변조를 탐지하는 시스템 및 이를 이용하여 무보정 탐지 서비스를 제공하는 방법
WO2021137946A1 (en) Forgery detection of face image
US20200218772A1 (en) Method and apparatus for dynamically identifying a user of an account for posting images
WO2021082562A1 (zh) 活体检测方法、装置、电子设备、存储介质及程序产品
CN110866466A (zh) 一种人脸识别方法、装置、存储介质和服务器
CN113642639B (zh) 活体检测方法、装置、设备和存储介质
JP2022133378A (ja) 顔生体検出方法、装置、電子機器、及び記憶媒体
CN115171199B (zh) 图像处理方法、装置及计算机设备、存储介质
WO2023071180A1 (zh) 真伪识别方法、装置、电子设备以及存储介质
Han et al. Two-stream neural networks for tampered face detection
WO2023272991A1 (zh) 一种数据处理方法、装置、计算机设备和存储介质
Einy et al. IoT Cloud‐Based Framework for Face Spoofing Detection with Deep Multicolor Feature Learning Model
Huang et al. Learnable Descriptive Convolutional Network for Face Anti-Spoofing.
Abavisani et al. A robust sparse representation based face recognition system for smartphones
CN115984977A (zh) 活体检测方法和系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22885090

Country of ref document: EP

Kind code of ref document: A1