A low-resolution real-time key information detection and protection method

Abstract

The application discloses a low-resolution real-time key information detection and protection method, which comprises the following steps: acquiring an image by using a network camera; using ShuffleNetV2 as a backbone network to extract sampling, using a Ge-Fusion FPN module for multi-scale feature fusion, and using an FCOS target detection head; simultaneously using a data distillation and a high-resolution pre-training method to improve the detection accuracy of a model under low resolution; and simultaneously using a chaotic encryption scheme to encrypt a detected key information region and fuse the key information region into an original image. The technical scheme of the application can improve the target detection speed and accuracy under low resolution, can be applied to the field of intelligent security and protection, and can improve the privacy performance of the network camera. The application improves the security by introducing a chaotic encryption method to protect the key information of the detected image in real time.

Description

Technical Field

[0001] This invention belongs to the field of intelligent security, specifically relating to a method for low-resolution real-time detection and protection of critical information. Background Technology

[0002] Network cameras, stored on cloud servers, offer users a more convenient experience. However, the interaction of large amounts of data with cloud servers can easily lead to data leaks. Therefore, better protecting user data security has become a critical issue. Traditional encryption systems, such as Data Encryption Standard (DES), Advanced Encryption Standard (AES), and Shamir Algorithm (RSA), are unsuitable for real-time image encryption due to drawbacks such as large data volumes, similar grayscale values, and high pixel correlation. Furthermore, these algorithms employ complex structures to increase randomness, thereby enhancing security. Consequently, traditional encryption algorithms require significantly more memory and computing power, which network cameras typically lack. Moreover, most existing algorithms encrypt the entire image. Sometimes, multimedia data contains a large amount of redundant information. Therefore, to improve encryption efficiency, avoid resource waste, and provide tiered privacy protection—allowing users with different permissions to access different levels of critical data—it is necessary to encrypt only critical areas of the image, balancing data availability and user privacy. Key information detection and encryption consist of two steps: detection and encryption. In target detection, current mainstream technologies like YOLOv7, YOLOv11 (Khanam R, Hussain M. YOLOv11: An overview of the key architectural enhancements[J]. arXiv preprint arXiv:2410.17725,2024.) perform detection at a 640x640 input resolution. However, in low-resolution network camera scenarios, real-time performance requires maintaining good detection results even with lower resolution image input. The Nanodet series offers good real-time performance, but accuracy still has room for improvement at a 320x320 input resolution. Regarding key information encryption, Wei Song (Song, W., Fu, C., Zheng, Y. et al. Protection of image ROI using chaos-based encryption and DCNN-based object detection. Neural Comput & Application) provides further research. 34,5743–5756 (2022). https: / / doi.org / 10.1007 / s00521-021-06725-w) proposed an encryption algorithm for key regions of interest (ROIs), which encrypts the auxiliary data of the key regions and embeds it inside the image. However, this method still needs to be performed on a 1080Ti desktop device, which consumes a lot of computing power and is difficult to implement the entire detection and encryption algorithm on low-computing-power embedded devices.Network security cameras require a higher frame rate for smooth operation. In the security field where network surveillance cameras are used, Aribilola et al. (Aribilola I, Asghar MN, Kanwal N, et al. SecureCam: Selective Detection and Encryption enabled Application for Dynamic Camera Surveillance Videos[J]. IEEE Transactions on Consumer Electronics, 2022.) proposed using H.264 encoding to encrypt images in the video stream based on pixel differences between consecutive frames. However, this encryption method still struggles to detect static critical information areas. Therefore, developing a method for real-time detection and protection of critical information on low-computing-power devices is of great significance in the field of intelligent security. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes a low-resolution real-time key information detection and protection method. To resolve the problems of low accuracy and slow speed in key information detection at low resolution, a lightweight low-resolution network is pre-trained using a higher resolution, achieving a significant improvement in accuracy. An LRRT-Det target detection network is proposed to acquire key information from network cameras. Simultaneously, a 2D-TEM chaotic sequence is proposed, and based on its characteristics, the two-stage ring diffusion algorithm is improved by performing more complex exchanges on the RGB three channels, encrypting the detection area, and encrypting the auxiliary data transmission of the embedded ROI region. Key information is transmitted by embedding the encrypted key information.

[0004] The present invention is achieved by at least one of the following technical solutions.

[0005] A method for low-resolution real-time detection and protection of critical information includes the following steps:

[0006] (1) Images obtained using a webcam;

[0007] (2) Input the image acquired by the webcam into the trained target detection model. The target detection model extracts and fuses features and outputs categories and detection box regions at different scales respectively.

[0008] (3) Use chaotic encryption to encrypt the contents of the detection box in real time, and embed the coordinate information of the encrypted area into the unencrypted area of ​​the image.

[0009] Furthermore, the target detection model is a low-resolution real-time target detection model LRRT-Det, which includes a feature extraction backbone network ShuffleNetV2, a generalized feature pyramid GeFusion-FPN network based on attention fusion, and a fully convolutional single-stage FCOS target detection head;

[0010] The feature extraction backbone network is used to extract multi-scale features of the image, GeFusion-FPN uses a feature selector to perform attention fusion on features of different scales, and the FCOS object detection head is used to output the region coordinates and detection category of the detection box in the final output.

[0011] Furthermore, the feature extraction backbone network uses ShuffleNetV2 to extract multi-scale features of the image, and inputs feature maps of three scales (1 / 8, 1 / 16, and 1 / 32) into GeFusion-FPN for feature fusion.

[0012] Furthermore, Ge-Fusion FPN performs attention fusion on features at different scales through a feature selector, and the fusion formula is as follows:

[0013] C = A × score [:,0] +B×score [:,1] ;

[0014] Where A and B are two features of the input Ge-Fusion FPN, and the score [:,0] The score represents the weight score corresponding to feature A. [:,1] Let B be the weight score corresponding to feature B, and C be the fusion feature output by Ge-Fusion FPN.

[0015] Furthermore, the training of the object detection model includes the following steps:

[0016] S1. Train the LRRT-Det model step by step using high-resolution images from the original dataset;

[0017] S2. Data distillation is performed using high-resolution images and the large object detection model Yolov11.

[0018] Further, in step S1, the LRRT-Det model is first trained at a scale of 416x416 for image input to obtain training parameters. The obtained training parameters are then used to initialize the object detection LRRT-Det model to guide the training of LRRT-Det with a 320x320 image input.

[0019] Data distillation in step S2: Use the Coco2017 and Widerface datasets to train a Yolov11 model at the input scale of 640x640 images. Use the output of the Yolov11 model as soft labels. The soft labels include not only the probability distribution of the target category, but also the relative relationship between each category, thus providing more learning information for the student model.

[0020] The teacher model outputs the class probability distribution and bounding box coordinates for each object through a softmax operation. The output z of the teacher model is... teacher Soft labels are obtained by smoothing at temperature T:

[0021]

[0022] Where, z teacher This is the output of the teacher model, where T is a temperature coefficient used to control the smoothness of the output; the student model minimizes the loss function L. total To achieve the best detection results, L total By utilizing both hard and soft labels from object detection, the losses from these two methods are weighted and combined, and the overall loss function is expressed as:

[0023]

[0024] Wherein, λ1, λ2, and λ3 are weighting coefficients that control the contribution ratios of hard labels, soft labels, and detection boxes, respectively. y represents the predicted class of the student model, and y represents the true class. It is the soft output of the teacher model. 'b' represents the coordinates of the detection box output by the student model, and 'b' represents the coordinates of the actual detection box. It is the hard-label classification loss, used to measure the difference between the student model's predicted class and the true label. It is the soft-label classification loss, used to measure the difference between the class predicted by the student model and the soft label generated by the teacher model. It is the bounding box regression loss, used to measure the difference between the bounding boxes predicted by the student model and the true bounding boxes.

[0025] Furthermore, in step (3), the chaotic sequence iterative equation used is:

[0026] x k+1 =mod(10·tanh(p1*y) k )·exp(x k +y k ),1);

[0027] y k+1 =mod(10·tanh(x) k)·exp(p2·(x k +y k ),1);

[0028] Where exp() represents the natural exponential function, tanh() represents the hyperbolic tangent function, p1 and p2 are control parameters with values ​​(0,1) set by the user, and y k x k It is the chaos value at the current k-th step, and mod(.) represents the modulo operation.

[0029] Further, in step (3), the encryption step is as follows: for each detection box of size w*h, where w and h are the number of rows and columns of the detection box region, firstly, different keys are associated according to the detection category, then the sequence is divided into a scrambling part and a diffusion part by using chaotic sequence iteration, the image is scrambled by using the method of elementary matrix row and column transformation, the pixels are diffused by using the XOR method, and finally the encrypted image replaces the detection box in the original image.

[0030] Furthermore, in step (3), information embedding is the embedding of information using a digital watermarking method.

[0031] The present invention provides a computer device comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the aforementioned low-resolution real-time critical information detection and protection method.

[0032] The present invention has the following advantages over the prior art:

[0033] 1. This invention introduces a weight allocation strategy into the feature fusion module (FPN module) in conventional object detection algorithms to become a generalized feature pyramid based on attention fusion (GeFusion-FPN), which enables better fusion of feature information at different scales and improves the accuracy of lightweight object detection under low-resolution image input in embedded environments.

[0034] 2. This invention improves security by introducing a chaotic encryption method to protect key information in the detected image in real time. Attached Figure Description

[0035] Figure 1 This is a flowchart of a low-resolution real-time critical information detection and protection method according to an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the training of the LRRT-Det target detection network according to an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the target detection network LRRT-Det structure according to an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the data fusion module in the GeFusion-FPN generalized feature pyramid based on attention fusion, according to an embodiment of the present invention.

[0039] Figure 5 This is a flowchart of the chaotic algorithm encryption process according to an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should be understood that the specific applications described are for explaining this application and are not intended to limit this application. The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0041] like Figure 1 As shown, this example provides a method for detecting and protecting critical information in low-resolution scenes, including the following steps:

[0042] (1) Images obtained using a webcam.

[0043] (2) Input the image acquired by the webcam into the trained target detection model LRRT-Det. The target detection model LRRT-Det uses the detection box to extract image features, fuses the extracted image features, and outputs the categories and detection box regions at different scales respectively.

[0044] The target detection model is a low-resolution real-time target detection model (LRRT-Det model), which includes the feature extraction backbone network ShfflenetV2, the attention-based generalized feature pyramid GeFusion-FPN (Generalized Aggregated Feature Fusion Module) network, and the FCOS (Fully Convolutional Single-Stage) target detection head; among which, GeFusion-FPN uses a feature selector to perform attention fusion on features of different scales to improve the accuracy of target detection.

[0045] As one example, the feature extraction backbone network uses ShuffleNetV2 to extract multi-scale features of the image, and inputs the multi-scale feature maps into a generalized feature pyramid network based on attention fusion for feature fusion.

[0046] Specifically, the feature extraction backbone network first samples the input 320x320 image by 8, 16, and 32 times to generate features at three scales: 40x40, 20x20, and 0x10, which are then input into the Ge-Fusion FPN network.

[0047] Ge-Fusion FPN introduces, for example Figure 3 The weight generator performs attention fusion on features at different scales to improve the accuracy of object detection. The key idea of ​​Ge-Fusion FPN is to further optimize feature representation by utilizing multi-scale features, combined with adaptive weight selection and attention mechanisms, thereby improving the performance of object detection models at different scales and in complex scenes.

[0048] The Ge-Fusion network, based on the original Feature Fusion Network (FPN), employs attention-based feature fusion to process the input of ShuffleNetV2 and the downsampling part of the FPN, such as... Figure 4 As shown, assuming the two input features are A and B, and the fusion feature is C, the weight generator module first concatenates the two types of data (A and B) to be fused along the channel dimension. Then, it extracts features through three convolutional layers. Finally, it obtains the weighted score for each channel using softmax. Based on the score weights, features A and B are fused using a weighted fusion formula:

[0049] C = A × score [:,0] +B×score [:,1] ;

[0050] Where A and B are two features of the input Ge-Fusion FPN, and the score [:,0] The score corresponds to the weighted score of input A. [:,1] Let B be the weight score corresponding to the input, and C be the fused feature output by Ge-Fusion FPN. Finally, bottom-up feature fusion is used to further improve the model's ability to learn details.

[0051] Finally, the fused data at the three scales is sent to the detection head, which is divided into two branches: classification and regression. Each branch first has two convolutional layers, followed by output convolutional layers for classification and regression, which output the categories and detection box regions at different scales respectively.

[0052] As one example, such as Figure 2 As shown, the training method for the object detection model LRRT-Det provided in this example includes the following steps:

[0053] 1) Use webcams to acquire key data from various scenarios as datasets, such as facial information and computer screen data, and label the data location and category information. Divide the training set and validation set in a 9:1 ratio.

[0054] 2) Train the Yolov11[3] model as a teacher network based on the training set to guide the training.

[0055] 3) Use a teacher network to perform data distillation and guide the training of the LRRT-Det model in high-resolution cases. This will improve the high accuracy of the LRRT-Det model at low resolution while ensuring that the LRRT-Det model has a low number of parameters and low computational cost.

[0056] 4) Use a high-resolution LRRT-Det model as the initial parameters, and simultaneously use a teacher network to train a low-resolution LRRT-Det model. For example... Figure 3 As shown in the figure, the structure of the final low-resolution detection model LRRT-Det is as follows.

[0057] In distillation, the teacher model obtains the class probability distribution and bounding box coordinates for each target by performing a normalized exponential function (Softmax) operation. To obtain soft labels, the teacher model's output z... teacher Smoothing by temperature T allows for richer information transfer between categories:

[0058]

[0059] Where, z teacher This is the soft label output by the teacher model; T is the temperature coefficient, which controls the smoothness of the output. It is the soft output after being weighted by the temperature coefficient and subjected to a Softmax operation. The student model LRRT-Det, on the other hand, minimizes L... total The loss function is used to achieve the best detection results. total By combining the hard labels (real labels) and soft labels (output of the teacher model) from object detection, and weighting the losses from both, the total loss function can be expressed as:

[0060]

[0061] Wherein, λ1, λ2, and λ3 are weighting coefficients that control the contribution ratios of hard labels, soft labels, and detection boxes, respectively. y represents the predicted class of the student model, and y represents the true class. It is the soft output of the teacher model. 'b' represents the coordinates of the detection box output by the student model, and 'b' represents the coordinates of the actual detection box. It is the hard-label classification loss, used to measure the difference between the class predicted by the student model and the true label. It is the soft label classification loss, used to measure the difference between the categories predicted by the student model and the soft labels generated by the teacher model. It is the bounding box regression loss, used to measure the difference between the bounding boxes predicted by the student model and the true bounding boxes.

[0062] (3) Use chaotic encryption method to encrypt the contents of the detection box in real time, and embed the coordinate information of the encrypted area into the unencrypted area of ​​the image using steganography.

[0063] After obtaining the key information image, such as Figure 5 The method uses chaotic encryption to protect the information security of images, specifically including a sequence generation part, a scrambling part, and a diffusion part:

[0064] 1) For each ROI region, first, based on the detected key information region of size w*h, where w and h are the number of rows and columns of the detection box region, respectively, then use chaotic sequence iteration. This process yields w*h+1000+w+h valid chaotic sequences. Then, the first 1000 chaotic sequences are removed to eliminate the influence of initial values ​​and increase disorder.

[0065] 2) In the scrambled part, first take the first w sequences x = {x} k}, where k = 1, 2, 3, ..., w, where x k (meaning) x k The values ​​of the first w chaotic sequences are sorted to generate a new descending matrix x′={x′ k Where k = 1, 2, 3, ..., w, x′ k It is x k The new sequence of elements after descending order, and at the same time determine {x k} to {x k The mapping of} forms a set T = {t} of mapped addresses. k}, t k For {x k The k-th element x in} k In {x′ k The sequence number in}, using {t k Generate a w*w elementary matrix P:

[0066]

[0067] Where P (m,n) Given the elements in matrix P at rows m and columns n, then using h sequences from (w+1) to (w+h), use the mapping {t} kThe method generates an h*h elementary matrix Q, and the RGB color image, i.e., the pixel matrix A of size w*h, is scrambled by performing elementary row and column transformations using matrix multiplication P*A*Q.

[0068] 3) In the diffusion part, first select w*h unused sequences to form a set X = {x} k1 Where k1 = 1, 2, 3, ..., w*h, and x k1 It represents the values ​​of w*h chaotic sequences required for the diffusion operation, where X is x. k1 The set that is composed of x, and for each x k1 There is X k1 =x k1 *2 24 , where X k1 It is for x k1 The amplified chaotic sequence values ​​are then extracted using a user key, and the last three sequences from the first 1000 sequences are used as s1, s2, and s3 for subsequent operations.

[0069] Pixel(1,1)=SR(1,1)+SG(1,1)*2 8 +SB(1,1)*2 16 ;

[0070]

[0071] SR(1,1)=mod(Pixel(1,1),2 8 );

[0072] SG(1,1)=mod(div(Pixel(1,1),2 8 ),2 8 );

[0073] SB(1,1)=mod(div(Pixel(1,1),2 16 ),2 8 );

[0074] Where SR(1,1) represents the R channel component of the image in the RGB color space at coordinate (1,1), SG(1,1) represents the G channel component of the image in the RGB color space at coordinate (1,1), SB(1,1) represents the B channel component of the image in the RGB color space at coordinate (1,1), and X[1] is the current chaotic sequence value. Pixel(1,1) is the pixel value after concatenating SR(1,1), SG(1,1), and SB(1,1), mod(div(.),2 8 ) represents taking 2 from div(.). 8 The modulus, i.e., div(.) divided by 2 8The remainder, div(Pixel(1,1),2 16 ) represents Pixel(1,1) divided by 2 16 The integer part of the quotient.

[0075] Additionally, the RGB space scrambling for the next iteration is determined by the previous pixel value, and an array D = {d i}, where i = (1, 2, 3), and its values ​​are shown below:

[0076]

[0077] Where Lx and Ly represent the x and y coordinates of the pixels in subsequent iterations, as shown below:

[0078]

[0079] Where i is the current iteration number, i = 2, 3, ..., w*h; mod((i-2),h) represents the remainder when i-2 is divided by h, and div((i-2),h) represents the integer part of the quotient when i-2 is divided by h.

[0080] Based on the above formula, D={d i}, Lx, Ly, and the remaining pixels are iterated using the following formula:

[0081]

[0082] SR(x,y)=mod(Pixel(x,y),2 8 );

[0083] SG(x,y)=mod(div(Pixel(x,y),2 8 ),2 8 );

[0084] SB(x,y)=mod(div(Pixel(x,y),2 16 ),2 8 );

[0085] Where X[n1] = 2, 3, ..., w*h, x = w, w-1..., 1, y = h, h-1, ... 1, X[n1] represents the expanded chaotic sequence X at position n1, SR(x,y) are the R channel components of the image in the RGB color space at coordinate (1,1), SG(x,y) are the G channel components of the image in the RGB color space at coordinate (1,1), SB(x,y) are the B channel components of the image in the RGB color space at coordinate (1,1), and Pixel(x,y) is the pixel value after concatenating SR(x,y), SG(x,y), and SB(x,y). Representing the XOR operation, d0, d1, d2 represent D = {d i The three components of}.

[0086] Finally, the image is encrypted, then the key information area is fused back into the original image, and digital watermarking technology is used to add the detection box coordinate information to the image, thus completing the encryption of the key information.

[0087] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for low-resolution real-time detection and protection of critical information, characterized in that, Includes the following steps: (1) Images acquired using a webcam; (2) Input the image acquired by the webcam into the trained target detection model. The target detection model extracts and fuses features and outputs categories and detection box regions at different scales respectively. The target detection model is the low-resolution real-time target detection model LRRT-Det, which includes the feature extraction backbone network ShuffleNetV2, the attention-based generalized feature pyramid GeFusion-FPN network, and the fully convolutional single-stage FCOS target detection head. Training an object detection model involves the following steps: S1. Train the LRRT-Det model step by step using high-resolution images from the original dataset; S2. Data distillation using high-resolution images and the large object detection model Yolov11: In step S1, the LRRT-Det model is first trained at a scale of 416x416 for image input to obtain training parameters. The obtained training parameters are then used to initialize the object detection LRRT-Det model to guide the training of LRRT-Det with 320x320 image input. Data distillation in step S2: Use the Coco2017 and Widerface datasets to train a Yolov11 model at the input scale of 640x640 images. Use the output of the Yolov11 model as soft labels. The soft labels include not only the probability distribution of the target category, but also the relative relationship between each category, thus providing more learning information for the student model. The teacher model outputs the class probability distribution and bounding box coordinates for each object through a softmax operation, while the student model minimizes the loss function. To achieve the best detection results, We utilize the hard and soft labels of object detection and combine the losses from both in a weighted manner. (3) Use chaotic encryption to encrypt the contents of the detection box in real time, and embed the coordinate information of the encrypted area into the unencrypted area of ​​the image. The encryption steps are as follows: for each detection box of size w*h, These are the number of rows and columns of the detection box region, respectively. First, different keys are associated according to the detection category. Then, the sequence is divided into a scrambling part and a diffusion part by using chaotic sequence iteration. The image is scrambled using elementary matrix row and column transformation. The pixels are diffused using the XOR method. Finally, the encrypted image replaces the detection box in the original image.

2. The method for low-resolution real-time detection and protection of critical information according to claim 1, characterized in that, The feature extraction backbone network is used to extract multi-scale features of the image. GeFusion-FPN uses a feature selector to perform attention fusion on features of different scales. The FCOS object detection head is used to output the region coordinates and detection category of the detection box in the final output.

3. The method for low-resolution real-time detection and protection of critical information according to claim 2, characterized in that, The feature extraction backbone network uses ShuffleNetV2 to extract multi-scale features of the image, and inputs feature maps of three scales (1 / 8, 1 / 16, and 1 / 32) into GeFusion-FPN for feature fusion.

4. The method for low-resolution real-time detection and protection of critical information according to claim 2, characterized in that, Ge-Fusion FPN uses a feature selector to perform attention fusion on features at different scales. The fusion formula is as follows: ； Where A and B are two features of the input Ge-Fusion FPN. The weight score corresponding to feature A. Let B be the weight score corresponding to feature B, and C be the fusion feature output by Ge-Fusion FPN.

5. The method for low-resolution real-time detection and protection of critical information according to claim 1, characterized in that, Output of the teacher model By temperature Smoothly obtain soft tags: ； in, It is the output of the teacher model. This is the temperature coefficient, used to control the smoothness of the output; the overall loss function is expressed as: ； in, These are weighting coefficients that control the contribution ratios of hard labels, soft labels, and detection boxes, respectively. The predicted category represents the student model. Represents the true category, It is the soft output of the teacher model. These are the coordinates of the detection box output by the student model. These are the actual bounding box coordinates. It is the hard-label classification loss, used to measure the difference between the student model's predicted class and the true label. It is the soft-label classification loss, used to measure the difference between the class predicted by the student model and the soft label generated by the teacher model. It is the bounding box regression loss, used to measure the difference between the bounding boxes predicted by the student model and the true bounding boxes.

6. The method for low-resolution real-time detection and protection of critical information according to claim 1, characterized in that, In step (3), the chaotic sequence iterative equation used is: Where exp() represents the natural exponential function, Represents the hyperbolic tangent function. The control parameter takes values ​​of (0,1), which are set by the user. , It is the current number k The chaos value of the step. (.) represents the modulo operation.

7. The method for low-resolution real-time detection and protection of critical information according to claim 1, characterized in that, In step (3), information embedding is the process of embedding information using digital watermarking.

8. A computer device, characterized in that, It includes: a memory and a processor, and a computer program stored in the memory, which, when executed on the processor, implements a low-resolution real-time critical information detection and protection method as described in any one of claims 1 to 7.

Images