AI Image Recognition-Based Security Monitoring Method and System

Abstract

The present invention relates to the field of image processing technology, and particularly to a security monitoring method and system based on AI image recognition. The method includes: obtaining a grayscale image of a target area; processing the grayscale image using an improved BM3D algorithm to obtain a denoised image; and performing security monitoring on the denoised image to obtain a warning result. That is, the solution of the present invention can improve the denoising effect of the image and facilitate the security monitoring of the image.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing. More specifically, the present invention relates to a security monitoring method and system based on AI image recognition. Background Art

[0002] In recent years, the progress of computer vision and artificial intelligence (AI) technologies has driven the rapid development of the security industry. The number of cameras in use is increasing, and AI intelligent applications based on cameras as the infrastructure have rapidly expanded and are widely used in various industries, such as production workshops, enterprise interiors, the financial field (monitoring of business halls, automated teller machines, and self-service banks, etc.), the telecommunications / electric power field (remote monitoring of switching rooms, wireless rooms, power rooms, etc., remote unattended monitoring of substations, power plants, etc.), the transportation field, community property management (security prevention in residential areas and offices, intelligent buildings, unmanned monitoring of parking lots, etc.), and the military security field, and so on.

[0003] Among them, cameras usually use computer vision and deep learning algorithms to implement image processing and visual recognition; image processing is an important step. Specifically, since the images collected in the monitoring scenario are easily affected by the external environment and the compression of the image quality during the transmission process, resulting in the loss of image details and the decline in quality, affecting the accuracy of subsequent recognition results; therefore, it is necessary to perform denoising preprocessing on the images before image recognition to improve the image quality and clarity for subsequent accurate recognition of the images.

[0004] Existing image denoising methods usually adopt filtering methods based on the spatial domain, wavelet decomposition in the transform domain, dictionary learning algorithms, or block-matching and 3D filtering (BM3D). Among them, the BM3D algorithm is a non-local denoising algorithm that uses the self-similarity and redundancy characteristics of images for denoising, such as Figure 1As shown in the figure, the algorithm is mainly divided into two stages: basic estimation and final estimation. Each stage includes three parts: block matching, collaborative filtering, and aggregation. The basic estimation provides weight parameters for the final estimation, and the final estimation filters the noise using the weights of the basic estimation. That is, the image I is first processed by the basic estimation to obtain a preliminary estimated image, and then the image I and the preliminary estimated image are processed through the final estimation stage to obtain the denoised image I1. Among them, block matching is to obtain several reference blocks by setting a sliding window. Each reference block searches in an appropriate surrounding area to find several blocks with the smallest difference (the difference between two blocks is usually quantified by the sum of squared errors algorithm), obtains the similar blocks of the reference block, and integrates these similar blocks into a three-dimensional matrix in any order. When determining the similar blocks, a threshold is usually set, and the blocks with a difference less than this threshold are used as similar blocks. At the same time, the maximum number of similar blocks is also set to determine the similar blocks of the reference block. Then, collaborative filtering processing is performed in the three-dimensional space, and the result is inverse-transformed and aggregated to form the denoised image. Therefore, this algorithm is an image denoising method that combines the spatial domain and the transform domain and is known as one of the best general image denoising algorithms with current denoising performance.

[0005] In the above process of block matching, the determination of similar blocks depends on the setting of the threshold. Therefore, whether the threshold is set too large or too small will lead to under-matching or over-matching in some areas, resulting in artifacts and ultimately poor image denoising effects. Summary of the Invention

[0006] The purpose of the present invention is to propose a security monitoring method and system based on AI image recognition to solve the problem of poor image denoising effect in the existing security monitoring process. For this purpose, the present invention provides solutions in the following two aspects.

[0007] In the first aspect, the present invention provides a security monitoring method based on AI image recognition, including:

[0008] Obtain the grayscale image of the target area;

[0009] Process the grayscale image using an improved BM3D algorithm to obtain a denoised image;

[0010] Perform security monitoring on the denoised image to obtain a warning result;

[0011] Among them, the method for adjusting the block matching in the basic estimation and / or final estimation of the improved BM3D algorithm is:

[0012] Obtain the reference blocks of the grayscale image and all candidate blocks of each reference block;

[0013] Take any reference block as the target reference block and calculate the similarity between the target reference block and any candidate block;

[0014] Calculate the distance between the center point of the target reference block and the central pixel point of any candidate block; take the product of the reciprocal of the distance and the first flatness degree of any candidate block as the similarity weight; the first flatness degree characterizes the difference change between the gray values of each pixel point in any candidate block and the average gray value of all pixel points; take the product of the similarity weight and the similarity as the weighted value; take the average value of the weighted values corresponding to all candidate blocks as the similarity threshold of the target reference block;

[0015] Take any candidate block with similarity greater than the similarity threshold as the similar block of the target reference block.

[0016] The beneficial effect is that: in the process of obtaining similar blocks in the basic estimation stage of the solution of the present invention, the similarity threshold is adaptively set, and a suitable similarity threshold can be determined for different reference blocks to realize the determination of the similarity of the reference blocks. Compared with the block matching grouping in the traditional BM3D algorithm, it avoids the problem that setting the same fixed threshold may lead to overmatching or under-matching in some regions, resulting in artifacts and thus poor denoising effect of the image.

[0017] In one embodiment, the first flatness degree is:

[0018] ; where is the first flatness degree of the candidate block , is the gray value of the i-th pixel point in the candidate block , is the gray average value of the candidate block , n is the total number of pixel points in the candidate block , is the maximum gray value in the candidate block , is the minimum gray value in the candidate block , is the set of gray values of all pixel points in the candidate block .

[0019] The beneficial effect is that: the first flatness degree of the candidate block is obtained through the gray difference between the gray values of all pixel points of the candidate block and its gray average value, and the change situation of all pixel points of the candidate block can be evaluated.

[0020] In one embodiment, the similarity is:

[0021] ; where is the similarity between the reference block C and the candidate block , is the reference feature value of the reference block C, is the candidate block 's candidate feature value, is the grayscale difference between the reference block C and the candidate block ; is the exponential function with the natural constant e as the base. The reference feature value is the product of the difference in the grayscale distribution interval and the peak factor corresponding to the grayscale histogram of the reference block C; the candidate feature value is the product of the difference in the grayscale distribution interval and the peak factor corresponding to the grayscale histogram of the candidate block .

[0022] The beneficial effect is that by using the difference in the grayscale distribution interval and the peak factor in the grayscale histograms of the reference block and the candidate block as their respective feature values, the similarity between the reference block and the candidate block can be accurately evaluated.

[0023] In one embodiment, the specific process of performing security monitoring on the denoised image is as follows:

[0024] Construct a target detection model, train the target detection model, and obtain a trained target detection model;

[0025] Input the denoised image into the trained target detection model to obtain a warning result, and perform a warning reminder according to the warning result, where the warning result includes dangerous targets and non-dangerous targets.

[0026] The beneficial effect is that by using the target detection model to perform target detection on the denoised image, the warning result can be accurately obtained.

[0027] In one embodiment, the target detection model is a Faster R-CNN model.

[0028] In one embodiment, the steps for obtaining the reference block of the grayscale image are as follows:

[0029] Use a preset window to traverse the entire grayscale image to obtain the pixel blocks corresponding to the window, and thus obtain all pixel blocks;

[0030] Obtain the second flatness degree of each pixel block, and use the pixel block corresponding to when the second flatness degree is less than the set threshold as the reference block; the second flatness degree characterizes the difference change between the grayscale values of each pixel point in the pixel block and the grayscale mean value of all pixel points.

[0031] The beneficial effect is that by obtaining the second flatness degree of each pixel block to screen the corresponding pixel blocks to determine the reference block, not only can the pixel blocks with large difference changes in the pixel blocks be screened out and used as the objects of concern, but also the subsequent computational complexity is reduced.

[0032] In one embodiment, the process of obtaining all candidate blocks of each reference block is as follows:

[0033] Taking any reference block as the center, set a search range and a target window of the same size as the reference block. Slide the target window to traverse the image area within the search range to obtain the candidate blocks corresponding to the target window, and obtain all candidate blocks of each reference block.

[0034] In a second aspect, a security monitoring system based on AI image recognition includes:

[0035] A processor;

[0036] A memory that stores computer instructions for security monitoring based on AI image recognition. When the computer instructions are run by the processor, the system executes the above-mentioned security monitoring method based on AI image recognition.

[0037] The beneficial effects of the present invention are as follows:

[0038] In the solution of the present invention, by setting a similarity threshold corresponding to each reference block during the process of obtaining similar blocks in the basic estimation stage, the similar blocks of each reference block can be determined more accurately, the error in the matching process is reduced, and thus the denoising effect of the image is improved, providing an accurate data basis for subsequent image monitoring.

[0039] Furthermore, the solution of the present invention also performs screening of reference blocks to obtain the reference blocks that need to be concerned, and only performs subsequent block matching, collaborative filtering, and aggregation on the screened reference blocks, reducing the amount of calculation and improving the calculation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0041] Figure 1 Schematically shows the basic flowchart of the existing BM3D algorithm;

[0042] Figure 2 Schematically shows the flowchart of the steps of the security monitoring method based on AI image recognition in this embodiment;

[0043] Figure 3 Schematically shows the structural block diagram of the security monitoring system based on AI image recognition in this embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the protection scope of the present invention.

[0045] The following will specifically describe the embodiments of the present invention in conjunction with the accompanying drawings.

[0046] The scenarios targeted by the present invention are mainly business halls, industrial parks, industrial production workshops, within enterprises, etc. It uses AI intelligence with cameras as the infrastructure to collect and identify images, and finally identifies whether there are potential hazards such as intrusion, theft, and fire. Among them, before image recognition, it is also necessary to perform filtering processing on the collected images. In the existing BM3D algorithm, in the matching process of similar blocks in the basic estimation and final estimation, a similarity threshold needs to be set manually to screen similar blocks. Since the content in different images is different, a fixed threshold may lead to overmatching or under-matching in some areas, resulting in artifacts. Therefore, the present invention provides a security monitoring method and system based on AI image recognition to improve the block matching process in the existing BM3D algorithm, reduce the artifacts generated in the matching process, and improve the image quality, which is beneficial to subsequent image recognition.

[0047] Specifically, taking a certain industrial park as an example, the security monitoring method based on AI image recognition of the present invention will be introduced.

[0048] As Figure 2 shown, the security monitoring method based on AI image recognition in this embodiment includes the following steps:

[0049] In step S1, a grayscale image of the target area is obtained. Specifically, a suitable location is selected in the industrial park for installing the camera, and the physical position and angle of the camera are adjusted to ensure that the field of view can cover the specified target area, and lens distortion correction is performed to ensure the geometric accuracy of the collected images.

[0050] In this embodiment, the image of the target area collected is grayscale processed to obtain a grayscale image of the target area. It should be noted that the sampling interval of the collected images is 5s, and the grayscale images of the collected images are denoised for subsequent image recognition.

[0051] In step S2, the improved BM3D algorithm is used to process the grayscale image to obtain a denoised image.

[0052] Among them, the method for adjusting the block matching in the basic estimation and / or final estimation of the improved BM3D algorithm is as follows:

[0053] First, obtain the reference blocks of the grayscale image and all candidate blocks of each reference block.

[0054] In this embodiment, a sliding window is set, and the entire grayscale image is traversed using the sliding window to obtain the pixel blocks corresponding to the sliding window, and then all pixel blocks are obtained; any pixel block can be used as a reference block.

[0055] In this embodiment, the size of the sliding window can be set to 5×5. Of course, the window size can also be adjusted according to the actual situation, but the window with an odd side length is required to ensure that there is a central pixel point in the window; among them, the sliding window slides on the entire image, and the step size of the sliding window during sliding is 5. Among them, each slide of the sliding window corresponds to a pixel block, and thus the entire grayscale image can be traversed to obtain multiple pixel blocks.

[0056] Furthermore, in order to reduce the computational amount, all pixel blocks can also be screened to obtain at least two reference blocks. Specifically, obtain the second flatness degree of each pixel block, and use the pixel block corresponding to the second flatness degree less than the set threshold as the reference block; the second flatness degree characterizes the change in the difference between the gray values of each pixel point in the pixel block and the average gray value of all pixel points.

[0057] Among them, the second flatness degree is: ; in the formula, is the second flatness degree of pixel block k, is the gray value of the i-th pixel point in pixel block k, is the average gray value of pixel block k, n is the total number of pixel points in pixel block k, is the maximum gray value in pixel block k, is the minimum gray value in pixel block k, is the set of gray values of all pixel points in pixel block k. When the size of the sliding window can be set to 5×5, the total number of pixel points n in pixel point k is 25.

[0058] In this embodiment, based on the cumulative sum of the differences between the gray value of each pixel point in the pixel block and the average gray value of the pixel block, the second flatness degree of the corresponding pixel block is reflected. If the cumulative sum of the differences is larger, the flatness degree of the pixel block is lower, the possibility of the corresponding pixel block containing important information (such as edge information) is greater, the necessity of filtering it to improve the image quality is greater, and the possibility of being screened as a reference block for filtering processing is greater.

[0059] The above set threshold can be set to 0.68, that is, when the flatness degree of pixel block k When (the experience value) is such that it can be considered that the pixel block k has a relatively low second flatness degree, the greater the possibility that the image content contains important information, and then it is screened as a reference block for filtering processing. Of course, the set threshold can also be set smaller to obtain a more suitable reference value.

[0060] In this embodiment, the pixel blocks are screened according to the second flatness degree to determine whether each pixel block is a reference block. It should be noted that the number of screened reference blocks can be at least two, which is specifically determined according to the actual screening situation.

[0061] In this embodiment, with any reference block as the center, a search range and a target window of the same size as the reference block are set. The target window is used to slide and traverse the image area within the search range to obtain the candidate blocks corresponding to the target window, and all candidate blocks of each reference block are obtained.

[0062] Among them, the search range set in this embodiment is 25×25. A target window of the same size as any reference block is set within its search range, and the step size is the side length of the target window. The target block corresponding to the target window within the search range is named a candidate block.

[0063] Exemplarily, if the size of the reference block is 5×5, then the size of the window is also 5×5. Since the reference block is randomly selected as any reference block, the size of other reference blocks is also 5×5.

[0064] It should be noted that when determining all candidate blocks in this embodiment, all candidate blocks may or may not include the reference block. In this embodiment, taking the case where all candidate blocks do not include the reference block as an example, then the subsequent similar blocks will definitely not include the reference block.

[0065] Secondly, taking any reference block as the target reference block, the similarity between the reference block and any candidate block is calculated.

[0066] The similarity in this embodiment can adopt cosine similarity, Euclidean distance or hash method. Since the cosine similarity, Euclidean distance or hash method are all existing technologies, the specific calculation process will not be elaborated here too much.

[0067] Of course, as other implementation manners, the similarity in this embodiment can also adopt the following calculation method, including the following steps:

[0068] Obtain the grayscale difference between the reference block and any candidate block within its search range; obtain the grayscale histogram of the reference block; based on the grayscale histogram, determine the corresponding grayscale distribution interval difference and peak coefficient; take the product of the grayscale distribution interval difference and the peak coefficient as the reference feature value of the reference block; obtain the grayscale histogram of the candidate block and its candidate feature value; obtain the similarity between the reference block and any candidate block according to the difference between the reference feature value and the candidate feature value and the grayscale difference.

[0069] The grayscale difference in the above steps is obtained by accumulating the grayscale differences of the corresponding position pixel points in the reference block and the candidate block, and taking the average of the accumulated differences as the grayscale difference between the reference block and the candidate block; specifically, for the reference block C and the candidate block The grayscale difference is expressed as: ; where is the grayscale value of the pixel point at the j-th position in the reference block C, is the grayscale value of the pixel point at the j-th position in the candidate block , and t is the serial number of the candidate block.

[0070] In this embodiment, the sizes of the reference block and the candidate block are both 5×5, and at this time, the number of pixel points in the reference block and the candidate block is both 25.

[0071] The grayscale distribution interval difference in the above steps is the difference between the maximum grayscale value and the minimum grayscale value in the grayscale histogram, and the peak coefficient is the ratio of the maximum grayscale frequency to the average grayscale frequency in the grayscale histogram. Since the grayscale distribution interval difference and the peak coefficient are prior arts, they will not be elaborated here in detail.

[0072] Exemplarily, the reference feature value of the grayscale histogram of the reference block C

[0073] .

[0074] Where is the maximum grayscale value in the reference block C, is the minimum grayscale value in the reference block C, is the maximum value of the grayscale frequency in the reference block C, is the average value of the grayscale frequency in the reference block C. represents the grayscale distribution interval difference of the reference block C, represents the peak coefficient of the grayscale histogram of the reference block C.

[0075] Among them, according to the above method of calculating the reference feature value, the feature value of the grayscale histogram of the candidate block , specifically: .

[0076] In the formula, is the maximum gray value in the candidate block , is the minimum gray value in the candidate block , is the maximum gray frequency in the candidate block , is the mean value of the gray frequency in the candidate block . represents the difference in the gray distribution interval of the candidate block , represents the peak factor of the gray histogram of the candidate block .

[0077] Furthermore, the similarity between the reference block C and the candidate block is:

[0078] .

[0079] In the formula, is the similarity between the reference block C and the candidate block , is the reference feature value of the reference block C, is the candidate feature value of the candidate block , is the gray difference between the reference block C and the candidate block , is the exponential function with the natural constant e as the base.

[0080] Among them, the smaller the difference in the histogram features and the gray difference between the reference block and the candidate block, the higher the similarity.

[0081] Then, calculate the distance between the center point of the target reference block and the center pixel point of any candidate block; multiply the reciprocal of the distance by the first flatness degree of any candidate block; the first flatness degree characterizes the change in the difference between the gray values of each pixel point in any candidate block and the gray mean value of all pixel points; multiply the similarity weight by the similarity as the weighted value; take the mean value of the weighted values corresponding to all candidate blocks as the similarity threshold of the target reference block.

[0082] Taking the reference block C as the target reference block as an example, the similarity weight in this embodiment is: ; among them, is the similarity weight between the reference block and the candidate block , is the reference block The position coordinates of the central pixel point, is the candidate block The position coordinates of the central pixel point, is the candidate block The first flatness degree.

[0083] Among them, the candidate block The first flatness degree is: ; In the formula, is the candidate block The gray value of the i-th pixel point in, is the candidate block The gray mean value, n is the candidate block The total number of pixel points in, is the candidate block The maximum gray value in, is the candidate block The minimum gray value in, is the candidate block The set of gray values of all pixel points in.

[0084] It should be noted that the reference block, pixel block and candidate block in this embodiment have the same size. Therefore, the total number of pixel points of the reference block, pixel block and candidate block is also the same, that is, all are n.

[0085] Among them, Is the distance between the center point of the reference block C and the position coordinates of the central pixel point of the candidate block

[0086] The similarity weight in the above is the product of the reciprocal of the Euclidean distance between the center point of the target reference block and the central pixel point of the candidate block and the flatness degree. When the distance is closer and the flatness degree of the candidate block is higher, the corresponding weight is larger.

[0087] Of course, as other implementation manners, in this embodiment, a rectangular coordinate system can also be established with the center point of the reference block as the origin, the direction of the abscissa of the center point as the x-axis, and the direction of the ordinate as the y-axis; calculate the similarity weight; at this time, only the distance from the position coordinates of the central pixel points of each candidate block within the search range of the reference block to the origin and the first flatness degree are required to obtain the corresponding similarity weight.

[0088] The similarity threshold in the above steps Is: ; In the formula, Is the similarity weight between the reference block And the candidate block , Is the similarity between the reference block And the candidate block , is the total number of candidate blocks.

[0089] It should be noted that each reference block in this embodiment corresponds to a similarity threshold, and the similarity thresholds of different reference blocks are different.

[0090] Finally, any candidate block with a similarity greater than the similarity threshold is used as the similar block of the target reference block.

[0091] In this embodiment, based on the similarity between the reference block and each candidate block and the set similarity threshold, some candidate blocks can be screened out as the similar blocks of the reference block for subsequent collaborative filtering processing.

[0092] Specifically, if , then the candidate block is used as the similar block of the reference block to obtain all the similar blocks of reference block C and perform subsequent collaborative filtering processing. is the similarity threshold. Therefore, in the process of screening similar blocks in this embodiment, an adaptive similarity threshold is set, that is, by setting the similarity thresholds corresponding to different reference blocks, the selected similar blocks are also different; avoiding the situation that when the similarity distribution of different reference blocks and their candidate blocks is different, using a fixed screening standard will also affect the subsequent filtering effect. That is, by adjusting the similarity measurement standard (similarity threshold) between the reference block and its candidate block, the algorithm robustness is improved, and at the same time, the filtering effect is also improved, which is convenient for subsequent AI recognition of images.

[0093] In the basic estimation stage, collaborative filtering and aggregation processing are performed on all the similar blocks corresponding to each obtained reference block; in the final estimation stage, collaborative filtering and aggregation processing are also performed on all the similar blocks corresponding to each obtained reference block, and finally a denoised grayscale image is obtained. Specifically, in the basic estimation process, when the similar blocks do not include the reference block, collaborative filtering combines the obtained reference block and its similar blocks to obtain a three-dimensional combination, and then performs collaborative transformation and filtering (3D-Transform) on it, and removes the energy of noise by setting a hard threshold (Hard-thresholding); aggregation processing is to integrate the filtering results of the combinations corresponding to each reference block (Aggregation). In the final estimation process, collaborative filtering uses the method of empirical Wiener shrinkage. According to the power spectrum of the collaborative transformation coefficient of the initial estimation image in the basic estimation process and the intensity of the noise, the collaborative transformation coefficient of the 3D block at the same position of the grayscale image is shrunk, and then the inverse transformation is performed on the shrunk coefficient, and then aggregation is performed to obtain the finally denoised image.

[0094] It should be noted that since only the block matching part is improved in this embodiment, while collaborative filtering and aggregation are not improved; therefore, since collaborative filtering and aggregation are prior arts, the process of how to perform collaborative filtering and aggregation after obtaining all similar blocks of the reference block will not be specifically introduced here, and details can be referred to the prior arts.

[0095] At step S3, perform security monitoring on the denoised image to obtain a warning result. Specifically, in this embodiment, by constructing a target detection model, training the target detection model to obtain a trained target detection model. Then input the denoised image into the trained target detection model to obtain a warning result, and perform a warning reminder based on the warning result.

[0096] The warning result in the above steps can be to classify the target area, such as dangerous targets and non - dangerous targets. When the warning result is a dangerous target, a warning reminder is required.

[0097] The target detection model in the above steps can be a YOLO (You Only Look Once) series model or a Faster R - CNN (Region with CNN feature) model.

[0098] The warning reminder in the above steps can issue a warning to relevant personnel through means such as sound, text message, email, etc. At the same time, it can be linked with other security systems, such as closing the gate, starting emergency equipment, etc.

[0099] In the process of security monitoring of the solution of the present invention, image processing is performed on the monitored image data to achieve denoising filtering of the image, obtain high - quality and high - definition images, and provide data support for subsequent image recognition.

[0100] The present invention also provides a security monitoring system based on AI image recognition. As Figure 3 shown, the security monitoring system includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the security monitoring method based on AI image recognition according to the above of the present invention is realized.

[0101] The security monitoring system also includes other components well - known to those skilled in the art, such as a communication bus and a communication interface. Their settings and functions are known in the art, so they will not be elaborated here.

[0102] In the present invention, the foregoing memory may be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, the computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium, such as, resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), and so on, or any other medium that can be used to store the required information and can be accessed by an application program, module, or both. Any such computer storage medium may be part of the device or accessible or connectable to the device. Any application or module described in the present invention may be implemented using computer-readable / executable instructions that can be stored or otherwise maintained by such a computer-readable medium.

[0103] In the description of this specification, "a plurality of" means at least two, such as two, three, or more, etc., unless otherwise specifically and clearly defined.

[0104] Although this specification has shown and described multiple embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will think of many changes, alterations, and alternative ways without departing from the spirit and scope of the present invention. It should be understood that various alternatives to the embodiments of the present invention described herein may be employed in the practice of the present invention.

Claims

1. A security monitoring method based on AI image recognition, characterized in that, Including: Obtain the grayscale image of the target area; Process the grayscale image by using an improved BM3D algorithm to obtain a denoised image; Conduct safety monitoring on the denoised image to obtain a warning result; Among them, the method for adjusting block matching in the basic estimation and / or final estimation of the improved BM3D algorithm is: Obtain the reference blocks of the grayscale image and all candidate blocks of each reference block; Take any reference block as the target reference block and calculate the similarity between the target reference block and any candidate block; Calculate the distance between the center point of the target reference block and the central pixel point of any candidate block; take the product of the reciprocal of the distance and the first flatness degree of any candidate block as the similarity weight; the first flatness degree characterizes the difference change between the gray values of each pixel point in any candidate block and the gray mean value of all pixel points; take the product of the similarity weight and the similarity as the weighted value; take the mean value of the weighted values corresponding to all candidate blocks as the similarity threshold of the target reference block; Take any candidate block with similarity greater than the similarity threshold as the similar block of the target reference block; The first flatness degree is: ； Wherein, is the first flatness of the candidate block , is the gray value of the i-th pixel in the candidate block , is the average gray value of the candidate block , n is the total number of pixels in the candidate block , is the maximum gray value in the candidate block , is the minimum gray value in the candidate block , is the set of gray values of all pixels in the candidate block .

2. The security monitoring method based on AI image recognition according to claim 1, wherein The similarity is: ； In the formula, is the similarity between the reference block C and the candidate block . is the reference feature value of the reference block C, is the candidate feature value of the candidate block , is the gray level difference between the reference block C and the candidate block , is an exponential function with the natural constant e as the base. The reference feature value is the product of the difference in the gray level distribution interval and the kurtosis coefficient corresponding to the gray level histogram of the reference block C; the candidate feature value is the product of the difference in the gray level distribution interval and the kurtosis coefficient corresponding to the gray level histogram of the candidate block .

3. The security monitoring method based on AI image recognition according to claim 1, wherein The specific process of conducting safety monitoring on the denoised image is: Construct a target detection model, train the target detection model to obtain a trained target detection model; Input the denoised image into the trained target detection model to obtain a warning result, and conduct warning reminders according to the warning result, where the warning result includes dangerous targets and non-dangerous targets.

4. The security monitoring method based on AI image recognition according to claim 3, wherein The target detection model is a Faster R-CNN model.

5. The security monitoring method based on AI image recognition according to claim 1, characterized in that The steps for obtaining the reference blocks of the grayscale image are: Use a preset window to traverse the entire grayscale image to obtain the pixel blocks corresponding to the window, and thus obtain all pixel blocks; Obtain the second flatness degree of each pixel block, and take the pixel block corresponding to the second flatness degree less than the set threshold as the reference block; The second flatness degree characterizes the difference change between the gray values of each pixel point in the pixel block and the gray mean value of all pixel points.

6. The security monitoring method based on AI image recognition according to claim 5, characterized in that The process of obtaining all candidate blocks of each reference block is: Take any reference block as the center, set a search range and a target window of the same size as the reference block, use the target window to slide and traverse the image area within the search range to obtain the candidate blocks corresponding to the target window, and obtain all candidate blocks of each reference block.

7. A security monitoring system based on AI image recognition, characterized in that, Including: A processor; A memory that stores computer instructions for safety monitoring based on AI image recognition. When the computer instructions are run by the processor, the system executes the safety monitoring method based on AI image recognition according to any one of claims 1-6.

Images