Monitoring screen area candid shooting detection method based on real-time visual identification

Abstract

The invention relates to the technical field of candid shooting detection, and discloses a monitoring screen area candid shooting detection method based on real-time visual identification, which comprises the following steps: determining a screen area, judging whether a suspected candid shooting behavior exists or not according to an overlapping result of a current image and a previous image of acquisition equipment, identifying a gesture action of the screen image, the method comprises the following steps: determining an object contour of a screen image according to an edge detection algorithm, searching an object contour value in a preset object contour table, obtaining a lens brightness value according to effective lens pixel points, outputting a contour weight factor of the object contour value and a brightness weight factor of the lens brightness value by adopting a candid shooting behavior model based on a gesture action, and comparing the candid photographing behavior value with a candid photographing behavior threshold value, and judging whether to send candid photographing early warning information to the signal shielding device or not according to a comparison result. The candid photographing behavior is comprehensively measured through the object contour value, the lens brightness value and the gesture action, and the candid photographing detection reliability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of voyeuristic camera detection, and in particular to a voyeuristic camera detection method for a monitoring screen area based on real-time visual recognition. Background Art

[0002] In recent years, with the continuous development of smart devices, screen hidden camera detection has faced new challenges.

[0003] Chinese patent application publication number CN117953574A discloses a screen anti-secret photography method and system based on visual recognition, including: when a user is in front of the screen of the display, real-time shooting is performed through a camera to obtain an original image; after denoising and enhancing the original image, an image to be detected is obtained; human body detection is performed on the image to be detected; when there is a human body area in the image to be checked, gesture recognition is performed on the human body area to determine whether there is a shooting behavior. It can be seen that simply using gestures without comprehensively measuring the characteristics of the object held will lead to certain limitations in identifying the shooting behavior.

[0004] Therefore, it is necessary to design a surveillance screen area candid detection method based on real-time visual recognition to solve the problems existing in current technology. Summary of the invention

[0005] In view of this, the present invention proposes a method for detecting candid photography in a surveillance screen area based on real-time visual recognition, aiming to solve the problem that simply using gestures without comprehensively measuring the characteristics of the object being held will lead to certain limitations in identifying the shooting behavior.

[0006] The present invention proposes a method for detecting hidden camera shots in a monitoring screen area based on real-time visual recognition, comprising: Determine a screen area, evenly deploy multiple acquisition devices in the screen area, and set the acquisition frequency of each acquisition device; Determine whether there is suspected candid photography behavior based on the overlap result of the current image and the previous image of the acquisition device. When the suspected candid photography behavior exists, obtain the current images of all acquisition devices and perform preprocessing, and determine the screen image of the screen area based on the result of the preprocessing; Recognize the gesture action of the screen image, determine the object contour of the screen image according to the edge detection algorithm, search the object contour value in the preset object contour table, determine the lens pixel point according to the object contour, analyze all the lens pixel points, determine the effective lens pixel point, and obtain the lens brightness value according to the effective lens pixel point; Based on the gesture action, a candid photography behavior model is used to output a contour weight factor of the object contour value and a brightness weight factor of the lens brightness value, a candid photography behavior value is determined according to the contour weight factor, the brightness weight factor, the object contour value and the lens brightness value, the candid photography behavior value is compared with a candid photography behavior threshold, and according to the comparison result, it is determined whether to issue a candid photography warning message to the signal blocker.

[0007] Furthermore, when judging whether there is suspected candid photography behavior based on the overlapping result of the current image and the previous image of the acquisition device, it includes: When all current images of all acquisition devices completely overlap with all previous images, it is determined that there is no suspected candid photography behavior in the screen area; When the current image and the previous image of a capture device do not completely overlap, it is determined that there is suspected candid photography behavior in the screen area.

[0008] Furthermore, when the current images of all acquisition devices are acquired and preprocessed, and the screen image of the screen area is determined according to the preprocessing result, it includes: The preprocessing includes image denoising and geometric correction, and extracting feature points from all current images after preprocessing, matching all extracted feature points, and determining the associated data between the current images; Based on the associated data, all current images are registered and merged to obtain a screen image to be processed in the screen area, and the screen image to be processed is image adjusted to determine the screen image of the screen area, wherein the image adjustment includes adjusting the image size and removing the splicing gap.

[0009] Further, when determining the object contour of the screen image according to the edge detection algorithm and searching the object contour value in the preset object contour table, it includes: The gradient of the screen image is calculated using the Sobel operator, non-maximum suppression is performed based on the Canny edge detection algorithm, and the edge detection result is optimized by a global optimal solution and a dynamic programming algorithm to extract the object contour of the screen image; Based on the object contour, a corresponding object contour value is determined in a preset object contour table, where the object contour table includes a mapping relationship between the object contour and the corresponding object contour value.

[0010] Furthermore, when analyzing all lens pixels to determine effective lens pixels and obtaining lens brightness values ​​according to the effective lens pixels, the following steps are included: All lens pixel points are changed into lens pixel position points, a lens pixel coordinate system is constructed from all lens pixel position points, all lens pixel position points are fitted, and a lens fitting curve is determined; The lens pixel position points that are not fitted are eliminated, the lens pixel position points on the lens fitting curve are used as valid lens pixel points, the brightness values ​​corresponding to all valid lens pixel points are obtained, and the average of all brightness values ​​is used as the lens brightness value.

[0011] Furthermore, when all lens pixel points are changed into lens pixel position points and a lens pixel coordinate system is constructed from all lens pixel position points, it includes: Taking the spatial position of the lens pixel point as the X-axis coordinate value of the lens pixel point, obtaining lens data of the lens pixel point, and establishing a lens data sequence based on the lens data; Acquire a qualified lens data sequence corresponding to the lens data sequence, and compare the lens data sequence with the qualified lens data sequence; When the lens data in the lens data sequence is greater than the qualified lens data in the qualified lens data sequence, the lens data is divided into a first lens data set; When the lens data in the lens data sequence is equal to the qualified lens data in the qualified lens data sequence, the lens data is divided into a second lens data set; When the lens data in the lens data sequence is less than the qualified lens data in the qualified lens data sequence, the lens data is divided into a third lens data set; The Y-axis coordinate value of the lens pixel point is calculated according to the first lens data set, the second lens data set and the third lens data set, and a lens pixel coordinate system is constructed according to the X-axis coordinate value and the Y-axis coordinate value.

[0012] Further, when calculating the Y-axis coordinate value of the lens pixel point according to the first lens data set, the second lens data set and the third lens data set, it includes: Counting a first number of shots in the first shot data set, a second number of shots in the second shot data set, and a third number of shots in the third shot data set; The Y-axis coordinate value of the lens pixel point is obtained by the following formula: ; Among them, R represents the Y-axis coordinate value of the lens pixel point, YA represents the mean of the lens data in the first lens data set, n represents the first number of lenses, Yi represents the i-th lens data in the first lens data set, Yg represents the qualified lens data corresponding to the i-th lens data in the first lens data set, YB represents the mean of the lens data in the third lens data set, m represents the third number of lenses, Yj represents the j-th lens data in the third lens data set, and Yh represents the qualified lens data corresponding to the j-th lens data in the third lens data set.

[0013] Furthermore, when the contour weight factor of the object contour value and the brightness weight factor of the lens brightness value are outputted using the candid photography behavior model based on the gesture action, it includes: Obtain a historical suspected candid photography data set, and divide the historical suspected candid photography data set into a training set and a test set; Preselect a convolutional neural network model, and substitute the training set into the convolutional neural network model for training; Verifying the trained convolutional neural network model according to the test set to determine the candid photography behavior model, and when the verification value is greater than or equal to the verification value after the previous training, using the currently trained convolutional neural network model as the candid photography behavior model; When the verification value is less than the verification value after the previous training, a grid search is used to find the hyperparameters of the convolutional neural network model, and the hyperparameters of the currently trained convolutional neural network model are replaced. The training set is clustered to determine a training sample set, and the training set is re-determined according to the training sample set. The re-determined training set is substituted into the convolutional neural network model with the replaced hyperparameters, and the training is continued until the candid photography behavior model is determined; The gesture action, the object contour value and the lens brightness value are substituted into the candid photography behavior model to obtain the contour weight factor and the brightness weight factor.

[0014] Furthermore, when clustering the training set to determine the training sample set, and re-determining the training set according to the training sample set, it includes: S30: determining an initial cluster number K according to the training set, and randomly selecting K points as initial cluster centers; S31: assigning each data point in the training set to the initial cluster center closest to it; S32: Calculate the mean of all data points in the cluster and use it as the new cluster center of the cluster; S33: Repeat S31 and S32 until the cluster center no longer changes or the predetermined number of iterations is reached, and the cluster finally determined is used as a training sample set; S34: interpolate between every two data points in the training sample set, determine a new data point and add it to the training set.

[0015] Further, when determining the candid photography behavior value according to the contour weight factor, the brightness weight factor, the object contour value and the lens brightness value, comparing the candid photography behavior value with the candid photography behavior threshold, and judging whether to send a candid photography warning message to the signal blocker according to the comparison result, the method includes: The sneak photography behavior value is obtained according to the following formula: ; Among them, Q represents the value of candid photography behavior, w1 represents the contour weight factor, S1 represents the object contour value, w2 represents the brightness weight factor, and S2 represents the lens brightness value; Pre-set the threshold Qa for candid photography behavior; When Q≥Qa, it is determined that a sneak shot warning message is sent to the signal blocker; When Q<Qa, it is determined that no sneak-photographing warning information is sent to the signal blocker.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: through image acquisition of multiple acquisition devices with real-time visual recognition, combined with analysis of object contours, calculation of lens brightness values ​​and capture of gesture actions, it is possible to accurately identify voyeuristic behavior, avoid relying on matching of single data or simple rules, and improve the reliability of voyeuristic behavior detection; multiple acquisition devices are evenly distributed to cover different angles of the screen area, thereby improving the ability to detect voyeuristic behavior in complex environments and avoiding the problem of a single acquisition device being blocked or limited by viewing angle; contour weight factors and brightness weight factors are determined based on the voyeuristic behavior model, and the object contour and lens brightness are comprehensively weighed, thereby improving the accuracy of calculating the voyeuristic behavior value; the voyeuristic behavior value is compared with the voyeuristic behavior threshold in real time, and early warning information can be sent to the signal blocker as soon as the voyeuristic behavior occurs, thereby preventing the voyeuristic device from working normally, effectively protecting personal privacy, and avoiding information leakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 A flowchart of a method for detecting hidden camera shots in a surveillance screen area based on real-time visual recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0019] In some embodiments of the present application, see Figure 1 As shown, a method for detecting hidden camera shots in a surveillance screen area based on real-time visual recognition comprises: S100: Determine a screen area, evenly deploy multiple acquisition devices in the screen area, and set the acquisition frequency of each acquisition device.

[0020] S200: Determine whether there is suspected candid photography based on the overlapping results of the current image and the previous image of the acquisition device. When there is suspected candid photography, obtain the current images of all acquisition devices and perform preprocessing, and determine the screen image of the screen area based on the preprocessing results.

[0021] S300: Identify gestures of the screen image, determine the object contour of the screen image according to the edge detection algorithm, and search for the object contour value in the preset object contour table, determine the lens pixel points according to the object contour, analyze all the lens pixel points, determine the effective lens pixel points, and obtain the lens brightness value according to the effective lens pixel points.

[0022] S400: Based on the gesture action, the candid photography behavior model is used to output the contour weight factor of the object contour value and the brightness weight factor of the lens brightness value, the candid photography behavior value is determined according to the contour weight factor, the brightness weight factor, the object contour value and the lens brightness value, the candid photography behavior value is compared with the candid photography behavior threshold, and it is determined whether to send a candid photography warning information to the signal blocker according to the comparison result.

[0023] Specifically, with the left edge of the screen as the center and the screen orientation as the reference, the center of 180 degrees is set as the screen area to ensure that all devices that may secretly take photos are within the detection range of the screen area, while avoiding detection blind spots and improving the accuracy of detection. Multiple acquisition devices are evenly deployed in the screen area, preferably 5 acquisition devices, and the acquisition devices are respectively captured at the 0°, 45°, 90°, 135° and 180° directions of the center. The acquisition device is a high-definition camera and the camera is equipped with an infrared sensor to enhance the image capture capability in a dark environment. The acquisition device captures images of the screen area at a set acquisition frequency of 30 frames per second. The acquisition frequency can be adjusted according to the actual screen area environment. According to the overlap result of the current image of the acquisition device and the image of the previous frame, the change of the current image is analyzed. If there is an abnormal overlap result, it is judged as suspected secretly taking photos. Once suspected secretly taking photos is found, the current images of all acquisition devices are obtained and preprocessed to accurately obtain the screen image to lay the foundation for subsequent analysis. As mentioned in the background technology, gesture actions are determined according to the trained gesture recognition model, and the object contour of the screen image is determined by the edge detection algorithm, so as to accurately judge the possible secretly photographing device (mobile phone, camera, etc.), extract the lens pixel points of the object contour, and determine the effective lens pixel points according to the fitting curve, so as to obtain the lens brightness value according to the effective lens pixel points, and then use the lens brightness characteristics of the secretly photographing device to improve the accuracy of secretly photographing detection, and find the object contour value in the preset object contour table. The object contour table is obtained according to the contours of multiple objects, such as (the size of a cylinder and the size of a cube, etc.). By combining the lens brightness value and the object contour value, the detection error caused by the mirror reflection of other interference objects can be avoided. For example, if only the lens brightness value is determined, the detection deviation may be caused by the mirror reflection of mirrors, watches, etc. If only the object contour value is determined, the detection deviation may be caused by similar objects (mobile phones and portable notebooks). Therefore, when performing secretly photographing detection on the screen area, it is necessary to combine the actual conditions of the two for analysis, and comprehensive detection cannot be achieved only for simple gesture actions.

[0024] It can be understood that, according to the determined gesture action, the candid photography behavior model is used to output the contour weight factor of the object contour value and the brightness weight factor of the lens brightness value, and the candid photography behavior value is determined according to the contour weight factor, the brightness weight factor, the object contour value and the lens brightness value. The brightness weight factor and the contour weight factor can balance the corresponding lens brightness value and object contour value, avoiding the data deviation caused by single data, thereby improving the accuracy of the candid photography behavior value, and comparing the candid photography behavior value with the candid photography behavior threshold to dynamically determine whether to send a candid photography warning message to the signal blocker. If a candid photography warning message is sent, the signal blocker can cut off the network interconnection between the communication device and the outside world, further preventing information leakage caused by candid photography, and improving the reliability of security protection of the screen area.

[0025] In some embodiments of the present application, when judging whether there is suspected voyeurism based on the overlapping results of the current image and the previous image of the acquisition device, it includes: when all current images of all acquisition devices completely overlap with all previous images, it is judged that there is no suspected voyeurism in the screen area; when the current image and the previous image of a acquisition device do not completely overlap, it is judged that there is suspected voyeurism in the screen area.

[0026] Specifically, when all current images of all acquisition devices completely overlap with all previous images, at this time, there is no movement of human beings or objects in the screen area, and it is determined that there is no suspected candid shooting behavior in the screen area. When the current image of a collection device does not completely overlap with the previous image, it indicates that there is a human being or object moving in the screen area, and it is determined that there is suspected candid shooting behavior in the screen area. Only when the current image changes will the suspected candid shooting behavior be dynamically determined, avoiding misjudgment caused by factors such as changes in ambient light and screen content, and improving the accuracy and reliability of candid shooting detection.

[0027] In some embodiments of the present application, when obtaining the current images of all acquisition devices and performing preprocessing, and determining the screen image of the screen area according to the results of the preprocessing, it includes: the preprocessing includes image denoising and geometric correction, and feature point extraction is performed on all the current images after preprocessing, all the extracted feature points are matched, and the correlation data between the current images is determined, and based on the correlation data, all the current images are aligned and merged to obtain the screen image to be processed of the screen area, and the screen image to be processed is adjusted to determine the screen image of the screen area, and the image adjustment includes adjusting the image size and removing the stitching gap.

[0028] Specifically, bilateral filtering is used for image denoising. Bilateral filtering retains the image details of the current image, while geometric correction corrects the distortion in all current images to ensure the alignment between the current images, thereby eliminating unnecessary information that affects subsequent processing and improving image quality. SIFT (Scale Invariant Feature Transform) is used for feature point extraction. SIFT can detect the same feature points at different scales. No matter how the scaling of the current image changes, SIFT can effectively identify feature points. According to all feature points extracted by SIFT, the RANSAC algorithm is used for matching to determine the associated data between the current images. The associated data includes the relative position and rotation relationship between the images. All current images are registered according to affine transformation to eliminate interference factors such as rotation and deformation caused by different shooting angles of the acquisition equipment, thereby improving the accuracy of the screen image to be processed. Multi-band fusion is used to merge all the registered current images to obtain the screen image to be processed. At this time, the screen image to be processed may have certain splicing gaps or size inconsistencies due to the processing of multiple algorithms. Through image adjustment, the splicing quality of the screen image can be improved, laying the foundation for subsequent analysis.

[0029] In some embodiments of the present application, when determining the object contour of a screen image according to an edge detection algorithm and searching for the object contour value in a preset object contour table, it includes: using the Sobel operator to calculate the gradient of the screen image, performing non-maximum suppression based on the Canny edge detection algorithm, and optimizing the edge detection result through a global optimal solution and a dynamic programming algorithm, extracting the object contour of the screen image, and determining the corresponding object contour value in a preset object contour table based on the object contour, wherein the object contour table includes a mapping relationship between the object contour and the corresponding object contour value.

[0030] Specifically, the Sobel operator is an edge detection operator that detects the edge of an image by calculating the gradient of the image. The Sobel operator convolves the screen image with two convolution kernels to detect changes in the horizontal and vertical directions respectively, obtains the gradient information of the screen image, and then identifies the edge of the screen image. The Canny edge detection algorithm suppresses the non-local maximum in the gradient direction and only retains the local maximum of the edge, thereby refining the edge and avoiding noise or false edges, thereby obtaining an accurate edge contour. The edge detection result is further adjusted through the global optimization algorithm (global optimization method in image segmentation) to make the edge contour smoother and more consistent. The global optimal solution takes into account the global information of the entire screen image, can suppress local errors and improve the accuracy of edge detection. The dynamic programming algorithm takes into account the relationship between adjacent pixels, eliminates errors caused by local noise or complex object shapes during edge detection, and improves the integrity and consistency of object contour extraction. The preset object contour table records the object contours of different objects and their corresponding object contour values. The object contour values ​​include length, angle, curvature, etc. By matching the extracted object contour with the object contour table, the object contour value in the screen image can be identified and determined, thereby improving the accuracy and efficiency of obtaining the object contour value.

[0031] In some embodiments of the present application, when analyzing all lens pixel points, determining effective lens pixel points, and obtaining lens brightness values ​​based on the effective lens pixel points, it includes: changing all lens pixel points into lens pixel position points, constructing a lens pixel coordinate system for all lens pixel position points, fitting all lens pixel position points, determining a lens fitting curve, eliminating lens pixel position points that are not fitted, taking the lens pixel position points on the lens fitting curve as effective lens pixel points, obtaining brightness values ​​corresponding to all effective lens pixel points, and taking the average of all brightness values ​​as the lens brightness value.

[0032] Specifically, in the actual screen image, although various image algorithms and preprocessing are used to avoid interference factors in the current image, the lens brightness value may still be deviated due to reflections from other objects, such as the optical reflections of mobile phone cameras and mirrors. The lens pixel points are determined according to the contours of the object, and all the lens pixel position points are fitted. The fitting process is implemented through OriginLab or Matlab, and the lens pixel position points that are not fitted are eliminated, eliminating some invalid lens pixel position points in the screen image due to optical reflection. The lens pixel coordinate system is expressed as a rectangular coordinate system, and the brightness values ​​corresponding to all valid lens pixel points are obtained, and the average of all brightness values ​​is used as the lens brightness value, which comprehensively measures the influence of all valid lens pixel points and avoids the detection error caused by the brightness value corresponding to a single valid lens pixel point.

[0033] In some embodiments of the present application, when all lens pixel points are changed into lens pixel position points and a lens pixel coordinate system is constructed from all lens pixel position points, the method includes: taking the spatial position of the lens pixel point as the X-axis coordinate value of the lens pixel point, obtaining lens data of the lens pixel point, and establishing a lens data sequence based on the lens data, obtaining a qualified lens data sequence corresponding to the lens data sequence, comparing the lens data sequence with the qualified lens data sequence, when the lens data in the lens data sequence is greater than the qualified lens data in the qualified lens data sequence, dividing the lens data into a first lens data set, when the lens data in the lens data sequence is equal to the qualified lens data in the qualified lens data sequence, dividing the lens data into a second lens data set, when the lens data in the lens data sequence is less than the qualified lens data in the qualified lens data sequence, dividing the lens data into a third lens data set, calculating the Y-axis coordinate value of the lens pixel point according to the first lens data set, the second lens data set and the third lens data set, and constructing a lens pixel coordinate system according to the X-axis coordinate value and the Y-axis coordinate value.

[0034] In some embodiments of the present application, when calculating the Y-axis coordinate value of the lens pixel point according to the first lens data set, the second lens data set and the third lens data set, it includes: counting the first number of lenses of the first lens data set, the second number of lenses of the second lens data set and the third number of lenses of the third lens data set, and the Y-axis coordinate value of the lens pixel point is obtained by the following formula: ; Among them, R represents the Y-axis coordinate value of the lens pixel point, YA represents the mean of the lens data in the first lens data set, n represents the first number of lenses, Yi represents the i-th lens data in the first lens data set, Yg represents the qualified lens data corresponding to the i-th lens data in the first lens data set, YB represents the mean of the lens data in the third lens data set, m represents the third number of lenses, Yj represents the j-th lens data in the third lens data set, and Yh represents the qualified lens data corresponding to the j-th lens data in the third lens data set.

[0035] Specifically, the spatial position of the lens pixel point is taken as the X-axis coordinate value of the lens pixel point. For example, only the spatial position of the lens pixel point on the x-axis (aligned with the screen), or the spatial position of the lens pixel point on the y-axis (perpendicular to the screen), or the spatial position of the lens pixel point on the z-axis (vertical to the ground) is considered. The lens data includes hue, saturation, etc. The qualified lens data in the qualified lens data series corresponds to the lens data one by one. By dynamically dividing all the lens data, the intelligence level of candid detection is improved, thereby improving the accuracy of the Y-axis coordinate value and ensuring the reliability of the lens pixel coordinate system.

[0036] In some embodiments of the present application, when a contour weight factor of an object contour value and a brightness weight factor of a lens brightness value are outputted using a candid behavior model based on gestures, the method includes: obtaining a historical suspected candid data set, dividing the historical suspected candid data set into a training set and a test set, pre-selecting a convolutional neural network model, substituting the training set into the convolutional neural network model for training, verifying the trained convolutional neural network model according to the test set, determining the candid behavior model, when the verification value is greater than or equal to the verification value after the previous training, using the currently trained convolutional neural network model as the candid behavior model, when the verification value is less than the verification value after the previous training, using a grid search to find the hyperparameters of the convolutional neural network model, and replacing the hyperparameters of the currently trained convolutional neural network model, clustering the training set to determine a training sample set, re-determining the training set according to the training sample set, substituting the re-determined training set into the convolutional neural network model with the replaced hyperparameters, continuing training until the candid behavior model is determined, substituting the gestures, object contour values ​​and lens brightness values ​​into the candid behavior model, and obtaining the contour weight factor and the brightness weight factor.

[0037] Specifically, the historical suspected candid data set includes various gestures from different periods, various object contour values, various lens brightness values, various contour weight factors, various historical brightness weight factors, and overall brightness change data of the screen area under different lighting conditions (ambient light intensity). The historical suspected candid data set is divided into a training set and a test set. Usually 50%-70% of the data is used as a training set, and the rest is used as a test set. The training set is used to train the convolutional neural network model, and the test set is used to verify the performance of the trained model. The convolutional neural network model is pre-selected. The convolutional neural network model contains multiple levels and different types of activation functions, aiming to capture the complex relationships in the data. The convolutional neural network model is trained using the data in the training set. During the training process, the convolutional neural network model learns the patterns and relationships in the data and iterates to improve its prediction or classification capabilities. After each training, the model is verified using the data in the test set. The verification includes accuracy and loss function values, etc., which are used to comprehensively measure the performance of the model.

[0038] It can be understood that if the verification value after the current training is greater than or equal to the verification value after the previous training, it means that the performance of the model has improved or remained relatively stable. At this time, the training is stopped and the currently trained convolutional neural network model is used as the voyeurism behavior model. If the verification value after the current training is less than the verification value after the previous training, it means that the performance of the model has declined. Grid search is used to find the hyperparameters of the convolutional neural network model, and the hyperparameters of the currently trained convolutional neural network model are replaced to change the learning rate of the model, which helps to find the most suitable network architecture and training strategy. The training process is repeated until the verification value is greater than or equal to the verification value after the previous training. The currently trained convolutional neural network model is used as the voyeurism behavior model, ensuring the data extraction and output capabilities of the model, so that the voyeurism behavior model can be judged from different angles, thereby improving the accuracy of the model output contour weight factor and brightness weight factor.

[0039] In some embodiments of the present application, when clustering the training set to determine the training sample set, and re-determining the training set according to the training sample set, it includes: S30: Determine the initial cluster number K according to the training set, and randomly select K points as initial cluster centers.

[0040] S31: Assign each data point in the training set to the initial cluster center closest to it.

[0041] S32: Calculate the mean of all data points in the cluster and use it as the new cluster center of the cluster.

[0042] S33: Repeat S31 and S32 until the cluster center no longer changes or the predetermined number of iterations is reached, and the clusters finally determined are used as training sample sets.

[0043] S34: interpolate between every two data points in the training sample set, determine new data points and add them to the training set.

[0044] Specifically, the k-means clustering algorithm can identify the potential patterns or structures of the data in the training set. The initial cluster number K can be adjusted according to the specific application scenario of the screen area. New data points are generated by interpolation, which effectively increases the data samples in the training set, allowing the model to be trained in a wide range of scenarios, thereby improving its adaptability to complex scenarios. The SMOTE algorithm is used for interpolation, the purpose of which is to generate new data points between the original data points. It also allows the model to be trained on the re-determined training set, avoiding overfitting of the model, allowing the model to make predictions on unseen data, thereby improving generalization ability. Clustering algorithms and interpolation algorithms help to deal with the problem of imbalanced categories in the training set. When there are fewer samples of certain categories in the training set, more samples can be generated by interpolation, thereby improving the balance of the categories. The re-determined training set effectively avoids the deviation of the accuracy caused by the model being biased towards a certain category during the training process, improves the accuracy of the contour weight factor and the brightness weight factor, and ensures the reliability of subsequent calculations.

[0045] In some embodiments of the present application, when determining the candid shooting behavior value according to the contour weight factor, the brightness weight factor, the object contour value and the lens brightness value, comparing the candid shooting behavior value with the candid shooting behavior threshold, and determining whether to send a candid shooting warning message to the signal blocker according to the comparison result, the method includes: The value of candid photography behavior is calculated according to the following formula: ; Among them, Q represents the voyeur behavior value, w1 represents the contour weight factor, S1 represents the object contour value, w2 represents the brightness weight factor, S2 represents the lens brightness value, and the voyeur behavior threshold Qa is set in advance. When Q≥Qa, it is determined to send a voyeur warning information to the signal blocker. When Q<Qa, it is determined not to send a voyeur warning information to the signal blocker.

[0046] Specifically, the contour weight factor and the brightness weight factor can comprehensively consider the characteristics of the object contour and the characteristics of the lens brightness, and avoid errors in the voyeuristic behavior value caused by a single data. For example, optical reflection and other similar voyeuristic objects are balanced by the contour weight factor and the brightness weight factor, thereby improving the accuracy of voyeuristic behavior detection. When the voyeuristic behavior value exceeds the voyeuristic behavior threshold, it is determined that there is voyeuristic behavior of the screen behavior. At this time, a voyeuristic warning message is sent to the signal blocker. The signal blocker receives the voyeuristic warning message and cuts off the mobile phone network and the wireless network, avoiding the communication equipment from communicating with the external network, and preventing information leakage and network propagation. In addition, a screen controller is added to the host or LED screen of the screen for voyeuristic detection, and the screen controller is used to change the screen refresh rate, insert black frames and PWM dimming. By utilizing the frequency range between the highest recognizable update frequency of the human eye and the shutter frequency of the camera device, the update frequency of the controllable display window of the screen is set so that the update frequency of the controllable display window of the screen is between the highest recognizable update frequency of the human eye and the shutter frequency of the camera device. By utilizing the frequency difference between the human eye, the candid camera and the controllable display window of the screen, the candid camera and the controllable display window of the screen are out of sync without affecting the viewing effect of the human eye, so that a complete picture cannot be obtained, thereby further preventing information leakage and effectively reducing candid photography behavior.

[0047] It is understandable that when the refresh rate is 275Hz, there is no visual abnormality when the human eye is watching, but the shutter speed (such as 50Hz / 60Hz or 1 / 120 second) of common mobile phones and camera devices is not synchronized, and a short black frame is inserted in the refresh cycle of the screen LED (such as inserting 1 black frame every 10 frames). The human eye ignores the black frame due to visual persistence, but the mobile phone and camera device will flicker when shooting due to the synchronization of the black frame and the shutter, thus avoiding information leakage. When dimming PWM, when the brightness of the screen LED is high-frequency PWM (such as 2000Hz), the human eye cannot perceive the flicker, but when the mobile phone and camera device shoot, the sampling interval is not synchronized with the PWM cycle, and the shooting result will have brightness fluctuations. When using low-frequency PWM (such as 200Hz), the human eye may perceive a certain amount of flicker at low brightness, and the mobile phone and camera device will enhance the flicker effect when shooting, resulting in poor image quality. It collects information through acquisition equipment and makes intelligent judgments, uses image recognition and gesture movements to detect voyeurism in real time, improves the accuracy and real-time nature of voyeurism detection, and uses the screen controller to change the screen refresh rate, insert black frames and PWM dimming to effectively prevent information leakage, thereby improving the security protection capabilities of the screen.

[0048] In summary, the beneficial effects of the present invention are as follows: through image acquisition of multiple acquisition devices with real-time visual recognition, combined with analysis of object contours, calculation of lens brightness values ​​and capture of gesture actions, it is possible to accurately identify voyeuristic behavior, avoid relying on matching of single data or simple rules, and improve the reliability of voyeuristic behavior detection; multiple acquisition devices are evenly distributed to cover different angles of the screen area, thereby improving the ability to detect voyeuristic behavior in complex environments and avoiding the problem of a single acquisition device being blocked or limited by viewing angle; contour weight factors and brightness weight factors are determined based on the voyeuristic behavior model, and the object contour and lens brightness are comprehensively weighed to improve the accuracy of calculating the voyeuristic behavior value; the voyeuristic behavior value is compared with the voyeuristic behavior threshold in real time, and an early warning message can be sent to the signal blocker as soon as the voyeuristic behavior occurs, thereby preventing the voyeuristic device from working normally, effectively protecting personal privacy, and avoiding information leakage.

[0049] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0050] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0051] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0052] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition, characterized in that: include: Determine a screen area, evenly deploy multiple acquisition devices in the screen area, and set the acquisition frequency of each acquisition device; Determine whether there is suspected candid photography behavior based on the overlap result of the current image and the previous image of the acquisition device. When the suspected candid photography behavior exists, obtain the current images of all acquisition devices and perform preprocessing, and determine the screen image of the screen area based on the result of the preprocessing; Recognize the gesture action of the screen image, determine the object contour of the screen image according to the edge detection algorithm, search the object contour value in the preset object contour table, determine the lens pixel point according to the object contour, analyze all the lens pixel points, determine the effective lens pixel point, and obtain the lens brightness value according to the effective lens pixel point; Based on the gesture action, a candid photography behavior model is used to output a contour weight factor of the object contour value and a brightness weight factor of the lens brightness value, a candid photography behavior value is determined according to the contour weight factor, the brightness weight factor, the object contour value and the lens brightness value, the candid photography behavior value is compared with a candid photography behavior threshold, and according to the comparison result, it is determined whether to issue a candid photography warning message to the signal blocker.

2. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 1 is characterized in that: When judging whether there is suspected voyeurism based on the overlapping results of the current image and the previous image of the acquisition device, it includes: When all current images of all acquisition devices completely overlap with all previous images, it is determined that there is no suspected candid photography behavior in the screen area; When the current image and the previous image of a capture device do not completely overlap, it is determined that there is suspected candid photography behavior in the screen area.

3. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 2 is characterized in that: When the current images of all acquisition devices are acquired and preprocessed, and the screen image of the screen area is determined according to the preprocessing result, it includes: The preprocessing includes image denoising and geometric correction, and extracting feature points from all current images after preprocessing, matching all extracted feature points, and determining the associated data between the current images; Based on the associated data, all current images are registered and merged to obtain a screen image to be processed in the screen area, and the screen image to be processed is image adjusted to determine the screen image of the screen area, wherein the image adjustment includes adjusting the image size and removing the splicing gap.

4. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 3 is characterized in that: When determining the object contour of the screen image according to the edge detection algorithm and searching the object contour value in the preset object contour table, it includes: The gradient of the screen image is calculated using the Sobel operator, non-maximum suppression is performed based on the Canny edge detection algorithm, and the edge detection result is optimized by a global optimal solution and a dynamic programming algorithm to extract the object contour of the screen image; Based on the object contour, a corresponding object contour value is determined in a preset object contour table, where the object contour table includes a mapping relationship between the object contour and the corresponding object contour value.

5. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 4 is characterized in that: When analyzing all lens pixels to determine effective lens pixels and obtaining lens brightness values ​​according to the effective lens pixels, the following steps are included: All lens pixel points are changed into lens pixel position points, a lens pixel coordinate system is constructed from all lens pixel position points, all lens pixel position points are fitted, and a lens fitting curve is determined; The lens pixel position points that are not fitted are eliminated, the lens pixel position points on the lens fitting curve are used as valid lens pixel points, the brightness values ​​corresponding to all valid lens pixel points are obtained, and the average of all brightness values ​​is used as the lens brightness value.

6. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 5 is characterized in that: When all lens pixel points are changed into lens pixel position points and all lens pixel position points are used to construct a lens pixel coordinate system, it includes: Taking the spatial position of the lens pixel point as the X-axis coordinate value of the lens pixel point, obtaining lens data of the lens pixel point, and establishing a lens data sequence based on the lens data; Acquire a qualified lens data sequence corresponding to the lens data sequence, and compare the lens data sequence with the qualified lens data sequence; When the lens data in the lens data sequence is greater than the qualified lens data in the qualified lens data sequence, the lens data is divided into a first lens data set; When the lens data in the lens data sequence is equal to the qualified lens data in the qualified lens data sequence, the lens data is divided into a second lens data set; When the lens data in the lens data sequence is less than the qualified lens data in the qualified lens data sequence, the lens data is divided into a third lens data set; The Y-axis coordinate value of the lens pixel point is calculated according to the first lens data set, the second lens data set and the third lens data set, and a lens pixel coordinate system is constructed according to the X-axis coordinate value and the Y-axis coordinate value.

7. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 6 is characterized in that: When calculating the Y-axis coordinate value of the lens pixel point according to the first lens data set, the second lens data set and the third lens data set, it includes: Counting a first number of shots in the first shot data set, a second number of shots in the second shot data set, and a third number of shots in the third shot data set; The Y-axis coordinate value of the lens pixel point is obtained by the following formula: ； Among them, R represents the Y-axis coordinate value of the lens pixel point, YA represents the mean of the lens data in the first lens data set, n represents the first number of lenses, Yi represents the i-th lens data in the first lens data set, Yg represents the qualified lens data corresponding to the i-th lens data in the first lens data set, YB represents the mean of the lens data in the third lens data set, m represents the third number of lenses, Yj represents the j-th lens data in the third lens data set, and Yh represents the qualified lens data corresponding to the j-th lens data in the third lens data set.

8. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 7 is characterized in that: When the contour weight factor of the object contour value and the brightness weight factor of the lens brightness value are outputted using the candid photography behavior model based on the gesture action, the method includes: Obtain a historical suspected candid photography data set, and divide the historical suspected candid photography data set into a training set and a test set; Preselect a convolutional neural network model, and substitute the training set into the convolutional neural network model for training; Verifying the trained convolutional neural network model according to the test set to determine the candid photography behavior model, and when the verification value is greater than or equal to the verification value after the previous training, using the currently trained convolutional neural network model as the candid photography behavior model; When the verification value is less than the verification value after the previous training, a grid search is used to find the hyperparameters of the convolutional neural network model, and the hyperparameters of the currently trained convolutional neural network model are replaced. The training set is clustered to determine a training sample set, and the training set is re-determined according to the training sample set. The re-determined training set is substituted into the convolutional neural network model with the replaced hyperparameters, and the training is continued until the candid photography behavior model is determined; The gesture action, the object contour value and the lens brightness value are substituted into the candid photography behavior model to obtain the contour weight factor and the brightness weight factor.

9. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 8, characterized in that: When clustering the training set to determine the training sample set, and re-determining the training set according to the training sample set, the method includes: S30: determining an initial cluster number K according to the training set, and randomly selecting K points as initial cluster centers; S31: assigning each data point in the training set to the initial cluster center closest to it; S32: Calculate the mean of all data points in the cluster and use it as the new cluster center of the cluster; S33: Repeat S31 and S32 until the cluster center no longer changes or the predetermined number of iterations is reached, and the cluster finally determined is used as a training sample set; S34: interpolate between every two data points in the training sample set, determine a new data point and add it to the training set.

10. The method for detecting hidden camera shots in monitoring screen areas based on real-time visual recognition according to claim 9, characterized in that: When determining the candid photography behavior value according to the contour weight factor, the brightness weight factor, the object contour value and the lens brightness value, comparing the candid photography behavior value with the candid photography behavior threshold, and judging whether to send a candid photography warning message to the signal blocker according to the comparison result, the method includes: The sneak photography behavior value is obtained according to the following formula: ； Among them, Q represents the value of candid photography behavior, w1 represents the contour weight factor, S1 represents the object contour value, w2 represents the brightness weight factor, and S2 represents the lens brightness value; Pre-set the threshold Qa for candid photography behavior; When Q≥Qa, it is determined that a sneak shot warning message is sent to the signal blocker; When Q<Qa, it is determined that no sneak-photographing warning information is sent to the signal blocker.

Images