Controlled cutter intelligent identification instrument and automatic judgment method

Abstract

The invention relates to the technical field of image recognition, and discloses a controlled cutter intelligent recognition instrument and an automatic judgment method, and the method comprises the steps: determining environment index data according to the environment detection data of the same type, and carrying out the qualification judgment of an image detection environment of a to-be-detected cutter according to all environment index data, the method comprises the following steps: determining a plurality of orientation detection images of a to-be-detected cutter, combining the orientation detection images to determine a to-be-detected image of the to-be-detected cutter, performing curve fitting on all cutter pixel points, determining pixel positions of the cutter pixel points which are not subjected to curve fitting, substituting the to-be-detected image into a cutter identification model to determine predicted pixel positions, and determining the to-be-detected cutter according to the predicted pixel positions. And judging whether to delete the cutter pixel points according to a comparison result, determining a cutter detection image of the to-be-detected cutter, comparing the cutter parameters with the standard cutter parameters, and determining a corresponding popup prompt according to the comparison result. The cutter is controlled through image recognition, and reliability and stability of detection are ensured.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a smart knife identification device and automatic determination method. Background Technology

[0002] Controlled knives, as highly dangerous prohibited items, pose a serious threat to public safety and social order. Currently, security checks on controlled knives mainly rely on traditional manual inspections and detection equipment. Traditional manual inspections involve visual checks and manual examinations by security personnel. However, during peak hours, handling large amounts of luggage and people can lead to fatigue and inattention among personnel, resulting in missed detections. Furthermore, manual judgment standards are subjective, with significant variations in accuracy when identifying concealed or modified controlled knives, and the inspection efficiency is low. Chinese Patent Publication No. CN110853019B discloses a method for detecting and identifying controlled knives during security checks, implemented according to the following steps: Step 1: Normalize image A to obtain image B; Step 2: Process image B using the SSD-ResNet101 model, feature fusion method, pooling, and ReLU function to obtain feature map C; Step 3: Calculate the default bounding box of feature map C using a scaling formula; use the JaccardOverlap matching strategy and model training to match the default bounding box with the target bounding box to obtain the predicted bounding box; Step 4: Use a non-maximum suppression algorithm to filter the final set of predicted bounding boxes to obtain the detection result. Therefore, it is evident that using X-rays to penetrate the object being inspected and generate an image of its internal structure to determine whether it is a cutting tool is a bulky and expensive device. Furthermore, it can only identify whether an object is a cutting tool, but it cannot determine the specific value of the tool or whether it is a controlled tool. In addition, environmental factors have a significant impact on the image, and the lack of an image recognition feedback mechanism makes the reliability of image recognition for cutting tools insufficient.

[0003] Therefore, it is necessary to design an intelligent identification device and automatic judgment method for controlled knives to solve the problems existing in the current technology. Summary of the Invention

[0004] In view of this, the present invention proposes an intelligent identification device and automatic judgment method for controlled knives, aiming to solve the problems of subjectivity in manual judgment standards, large differences in the accuracy of identification of controlled knives with concealed appearance and modification, low inspection efficiency, large size and high cost of X-ray equipment, which can only identify whether it is a knife, but cannot determine the specific value of the knife or whether the knife is a controlled knife, the significant influence of environmental factors on the image, the lack of image recognition feedback mechanism, and insufficient reliability of image recognition for knives.

[0005] In one aspect, the present invention proposes an automatic determination method for a smart knife identification device, comprising: The tool to be inspected is identified, and several environmental detection data of the tool to be inspected are acquired based on the acquisition interval. Environmental index data are determined based on environmental detection data of the same type. The passability of the image detection environment of the tool to be inspected is judged based on all environmental index data. If the image detection environment of the tool to be detected is qualified, then several orientation detection images of the tool to be detected are determined, and the several orientation detection images are merged to determine the image to be detected of the tool to be detected. Extract the knife pixel points of the image to be detected, and perform curve fitting on all knife pixel points to determine the pixel positions of the knife pixel points that are not curve fitted. Substitute the image to be detected into the knife recognition model to determine the predicted pixel positions, and compare the pixel positions with the predicted pixel positions. Based on the comparison results, determine whether to delete the knife pixel points to determine the knife detection image of the knife to be detected. The tool detection image is analyzed to determine the tool parameters of the tool to be detected, and the tool parameters are compared with the standard tool parameters. Based on the comparison results, a corresponding pop-up prompt is generated.

[0006] Furthermore, when acquiring several environmental detection data points of the tool under test based on the acquisition interval, and determining environmental indicator data based on environmental detection data of the same type, the process includes: Obtain all environmental monitoring data of the same type, and determine the average value of all environmental monitoring data of the same type as the environmental indicator data.

[0007] Furthermore, when determining the passability of the image detection environment for the tool under test based on all environmental index data, the following steps are included: Determine the range of environmental indicator data corresponding to each environmental indicator data, and compare each environmental indicator data with its corresponding range of environmental indicator data; If each environmental indicator data is within the corresponding environmental indicator data range, then the image detection environment of the tool to be tested is deemed qualified. If one or more environmental indicator data are within the corresponding environmental indicator data range, then the image detection environment of the tool to be tested is determined to be unqualified.

[0008] Furthermore, when merging several orientation detection images to determine the image to be detected for the cutting tool, the process includes: Several orientation detection images are preprocessed, including image denoising and color equalization. Feature points are extracted from the preprocessed orientation detection images using the BRISK algorithm, and the extracted feature points are matched using the FLANN algorithm to determine the spatial transformation relationship between the preprocessed orientation detection images. The spatial transformation relationship includes relative position and rotation relationship. The preprocessed orientation detection images are registered based on the spatial transformation relationship, and the registered orientation detection images are merged using the pyramid fusion algorithm. The merged orientation detection images are then stitched together to determine the image to be detected for the tool to be detected. The stitching process includes removing stitching gaps and sharpening image edges.

[0009] Furthermore, when extracting the tool pixels from the image to be detected and performing curve fitting on all tool pixels to determine the pixel positions of tool pixels that are not curve-fitted, the process includes: The preset deployment time is used as the X-axis coordinate value of each tool pixel, and the tool pixel value of each tool pixel is used as the Y-axis coordinate value. A tool coordinate system is established based on the X-axis coordinate value and the Y-axis coordinate value, and each tool pixel is converted into a tool pixel coordinate point according to the X-axis coordinate value and the Y-axis coordinate value. Curve fitting is performed on all tool pixel coordinate points to determine the tool pixel fitting curve, and the tool pixel position of the unfitted tool pixel coordinate points is determined based on the tool pixel fitting curve.

[0010] Furthermore, when substituting the image to be detected into the tool recognition model to determine the predicted pixel position, the process includes: Obtain a tool sample dataset and divide the tool sample dataset into a training set and a test set; The image network model is trained using the training set and tested using the test set. Finally, the input is determined to be the image to be detected, and the output is the tool recognition model with the predicted pixel position. The image network model includes a BP neural network model or an RBF neural network model.

[0011] Furthermore, when comparing the pixel position with the predicted pixel position, and determining whether to delete the tool pixel based on the comparison result to determine the tool detection image of the tool to be detected, the process includes: The tool pixel position is compared with the predicted pixel position; If the tool pixel position is consistent with the predicted pixel position, the corresponding tool pixel is deleted, and the tool detection image of the tool to be detected is determined based on the deletion result. If the tool pixel position and the predicted pixel position are inconsistent, the image to be detected is determined as the tool detection image of the tool to be detected.

[0012] Furthermore, when analyzing the tool detection image to determine the tool parameters of the tool to be detected, the process includes: The tool detection image is analyzed based on the contour detection algorithm to determine the tool parameters of the tool to be detected. The tool parameters include the tip, cutting edge, handle, tail, blood groove, blade, guard, and length.

[0013] Furthermore, when comparing the tool parameters with standard tool parameters and determining the corresponding pop-up prompt based on the comparison result, the following steps are included: The tool parameters are compared with the standard tool parameters; If the data in the tool parameters is consistent with the standard data in the standard tool parameters, the tool to be tested is determined to be a controlled tool and a red pop-up window is issued; otherwise, the tool to be tested is determined to be a non-controlled tool and a green pop-up window is issued.

[0014] Compared with existing technologies, the advantages of this invention are as follows: In the initial stage of detection, by acquiring environmental detection data of the tool under test based on the acquisition interval, determining environmental index data, and judging the passability of the image detection environment of the tool under test, adverse detection environments are screened in advance, avoiding interference from environmental factors on image acquisition quality. This avoids the risk of insufficient image recognition reliability due to the lack of an environmental adaptation mechanism, providing a stable environmental foundation for image acquisition. Acquiring several orientation detection images and merging them into the image under test integrates tool information from multiple perspectives, compensating for the limitations of a single image's perspective and incomplete features. This allows the image under test to fully present various structural details of the tool under test. After extracting the tool pixels from the image under test, the positions of unfitted pixels are determined through curve fitting. Based on the knife recognition model, the predicted pixel positions are determined, and redundant pixels are removed to eliminate noise, background residue, and other invalid information in the image, resulting in a knife detection image that accurately reflects the true structure of the knife to be detected. This solves the risk of feature distortion caused by pixel mixing and improves the accuracy of parameter analysis. By comparing the knife parameters with standard knife parameters, it not only addresses the shortcomings of existing methods in determining the specific parameters of the knife, but also accurately determines whether the knife to be detected is a controlled knife. This replaces the subjectivity and experience dependence of manual judgment, avoiding missed detections and misjudgments. At the same time, red and green pop-up prompts allow security personnel to quickly obtain the judgment results and take corresponding actions, balancing the rigor and smoothness of security inspection work, and providing reliable support for the prevention and control of controlled knives in the field of public safety.

[0015] On the other hand, this application also provides a smart identification device for controlled knives, for using the above-mentioned automatic determination method of the smart identification device for controlled knives, including: Base, display screen, support columns, crossbeams, and support blocks; One end of the support column is fixedly connected to the base, and the other end of the support column away from the base is fixedly connected to the crossbeam; One end of the support block is fixedly connected to the support column, and the other end of the support block away from the support column is fixedly connected to the display screen. A camera is installed on the crossbeam.

[0016] It is understandable that the aforementioned intelligent identification device and automatic judgment method for controlled knives have the same beneficial effects, which will not be elaborated here. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the intelligent identification device and automatic judgment method for controlled knives provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the intelligent identification device for controlled knives provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure along direction A provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the crossbeam provided in an embodiment of the present invention.

[0019] The components include: 1. base; 2. display screen; 3. support column; 4. crossbeam; 5. camera; and 6. support block. Detailed Implementation

[0020] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] See Figure 1 As shown in some embodiments of this application, an automatic determination method for a smart knife identification device includes: S100: Determine the tool to be inspected, acquire several environmental detection data of the tool to be inspected based on the acquisition interval, determine environmental index data based on environmental detection data of the same type, and determine the passability of the image detection environment of the tool to be inspected based on all environmental index data.

[0023] S200: If the image detection environment of the tool to be inspected is qualified, then determine several orientation detection images of the tool to be inspected, and merge the several orientation detection images to determine the image to be inspected of the tool.

[0024] S300: Extract the tool pixels in the image to be detected, perform curve fitting on all tool pixels, determine the pixel position of the tool pixels that are not curve fitted, substitute the image to be detected into the tool recognition model to determine the predicted pixel position, compare the pixel position with the predicted pixel position, and determine whether to delete the tool pixels based on the comparison result, and determine the tool detection image of the tool to be detected.

[0025] S400: Analyzes the tool inspection image, determines the tool parameters of the tool to be inspected, compares the tool parameters with the standard tool parameters, and determines the corresponding pop-up prompt based on the comparison result.

[0026] Specifically, after identifying the tool to be inspected and clearly defining the object, several environmental detection data points of the environment in which the tool is located are acquired based on a collection interval of once every five seconds, for a total of one minute. The specific collection interval and corresponding time can be dynamically adjusted according to the flow of people. The collected environmental detection data covers environmental factors that affect the quality of image acquisition, including light intensity, ambient humidity, ambient temperature, and ambient air pressure. The corresponding acquisition equipment can be miniature light sensors, humidity sensors, temperature sensors, and air pressure sensors, which are deployed in the environment in which the tool is located. Within one minute, twelve data points of each type of light intensity, ambient humidity, ambient temperature, and ambient air pressure are acquired. These data together constitute the environmental detection data. Non-single-collection determination of environmental detection data avoids the randomness of single-time data acquisition, ensuring that the captured data can truly reflect the patterns of environmental changes. The acquired environmental detection data are classified by type, and environmental indicator data are determined by integrating environmental detection data of the same type. That is, environmental detection data of the same type are summarized and analyzed to form a comprehensive indicator that can accurately reflect the true state of the current detection environment. Based on all environmental indicator data, the image detection environment of the tool under test is assessed for compliance. This eliminates the impact of adverse environments such as insufficient lighting and excessive background interference on image acquisition, avoiding the risk of image blurring and feature loss due to environmental factors, and ensuring the accuracy of subsequent image recognition of the tool under test. If the image detection environment of the tool under test is qualified, several azimuth detection images of the tool are determined. Since images from a single angle cannot fully present all structural features of the tool under test and have blind spots, which may lead to occlusion or feature loss of parts of the tool under test, images of the tool under test are acquired from multiple different angles and azimuths, including oblique angles and frontal images of the tool under test. These azimuth detection images are merged and processed by image fusion to integrate information from each azimuth detection image, supplementing the structural features missing in a single image, and finally determining the image of the tool under test. This achieves comprehensive coverage of the structural information of the tool under test. Compared with the method of relying on only a single image processing, the image under test can more completely capture the shape contour and structural details of the tool, providing a comprehensive and accurate image foundation for subsequent pixel extraction and parameter analysis.

[0027] Understandably, the process involves extracting tool pixels from the merged image to be detected, accurately capturing all pixels belonging to the tool, and then performing curve fitting on all extracted tool pixels. Curve fitting delineates the overall contour curve of the tool based on its shape characteristics. Based on the continuous structural characteristics of the tool, a smooth curve reflecting its true shape is fitted. At this point, tool pixels not fitted by the curve are likely redundant pixels caused by image noise, residual background pixels, or acquisition errors. To accurately distinguish between valid and noisy pixels, the image to be detected is fed into a trained tool recognition model. Through training with a large number of tool samples, the model can accurately identify the distribution pattern of valid pixels in the tool to be detected, and then output the predicted pixel positions that the tool should have, i.e., the pixel area range that conforms to the normal structural characteristics of the tool. The actual pixel positions of the knife pixels that are not fitted by the curve are compared with the predicted pixel positions output by the model. Through this comparison and elimination, the knife detection image of the knife to be detected is finally determined. This achieves precise purification of the knife pixels, eliminating interference factors such as image noise and background residue, and ensuring that the knife detection image contains only effective pixels that reflect the real structure of the knife. The knife parameters of the knife to be detected are extracted through image measurement and feature recognition. These parameters cover the core indicators for judging controlled knives, including specific parameters that accurately reflect the knife attributes, such as the blade tip, blade length, blade width, and handle length. The obtained knife parameters are comprehensively compared with the standard knife parameters, which are determined according to the "Standard for the Identification of Controlled Knives". It covers controlled knives such as daggers, triangular scrapers, and spring knives (switchblades). By comparing them one by one, it is possible to accurately determine whether the knife to be detected meets the identification standard for controlled knives. Based on the comparison results, corresponding pop-up prompts are generated, and the comparison process is displayed simultaneously for security personnel to review and refer to. This achieves automated and accurate judgment of controlled knives, avoiding the subjectivity and reliance on experience in manual judgment, and making up for the shortcomings of existing intelligent detection technologies that cannot determine the controlled attributes of knives. The pop-up prompts can intuitively present the judgment results, improving security inspection efficiency while ensuring the reliability and accuracy of the judgment results.

[0028] In some embodiments of this application, when acquiring several environmental detection data of the tool to be tested based on the acquisition interval, and determining environmental indicator data based on environmental detection data of the same type, the method includes: acquiring all environmental detection data of the same type, and determining the average value of all environmental detection data of the same type as the environmental indicator data.

[0029] In some embodiments of this application, when determining the passability of the image detection environment of the tool to be tested based on all environmental index data, the process includes: determining the range of environmental index data corresponding to each environmental index data, comparing each environmental index data with the corresponding range of environmental index data, and if each environmental index data is within the range of the corresponding environmental index data, then the image detection environment of the tool to be tested is deemed to be qualified; if one or more environmental index data are within the range of the corresponding environmental index data, then the image detection environment of the tool to be tested is deemed to be unqualified.

[0030] Specifically, after identifying the tool to be inspected as the object of identification, several environmental detection data points of the detection environment in which the tool is located are continuously acquired at preset acquisition intervals. These data are classified by type, covering environmental factors such as light and humidity that affect the quality of image acquisition. All environmental detection data of the same type are extracted, and the mean of this type of data is determined as the corresponding environmental index data. Since the data acquired in a single acquisition may be abnormal due to instantaneous interference, the mean can offset the random fluctuations and integrate the overall trend of the data of this type, eliminating the influence of abnormal data. This ensures that the environmental index data truly and objectively reflects the actual state of the corresponding environmental factors, avoiding misjudgment of the environmental state due to a single abnormal data point, and providing an accurate and reliable basis for subsequent environmental compliance determination. Environmental data includes four types: light intensity, ambient humidity, ambient temperature, and ambient air pressure. For image recognition cameras, the degree to which these four types affect image recognition accuracy is determined through factory instructions or experimental simulations. This allows for the determination of the data range for each type of environmental indicator. For example, through experimental simulations, three environmental factors are fixed, while only the fluctuation range of one factor is changed. Image recognition accuracy data under different fluctuation states is recorded simultaneously, thus accurately pinpointing the degree of influence and critical threshold of each factor on recognition accuracy. Light intensity has the most direct impact on image clarity and feature contrast; excessive brightness or darkness can lead to the loss of pixel features. Simulations can determine the stable range of light intensity, thereby determining the range of each environmental indicator. The data is compared one by one with the corresponding environmental indicator data range. If all environmental indicator data fall within their respective ranges, it means that the current image detection environment can accurately identify the tool to be detected. Otherwise, if any factor fails to meet the standard, the acquired tool image will be blurry, feature lost or distorted, which will affect the accuracy of subsequent pixel extraction and analysis. In this case, the image detection environment is deemed unqualified. This not only avoids the subjective bias of manual judgment, but also ensures that the subsequent detection stage is only entered when all environmental conditions meet the image acquisition requirements. This reduces the risk of misjudgment and omission due to environmental problems and provides a guarantee for the stability and accuracy of the entire automatic judgment process for controlled tools.

[0031] In some embodiments of this application, when merging several orientation detection images to determine the image to be detected for the tool, the process includes: preprocessing the several orientation detection images, including image denoising and color equalization; extracting feature points from the preprocessed orientation detection images based on the BRISK algorithm and matching the extracted feature points based on the FLANN algorithm; determining the spatial transformation relationship between the preprocessed orientation detection images, including relative position and rotation relationships; registering the preprocessed orientation detection images based on the spatial transformation relationship; merging the registered orientation detection images using a pyramid fusion algorithm; and stitching the merged orientation detection images to determine the image to be detected for the tool, wherein the stitching process includes removing stitching gaps and sharpening image edges.

[0032] Specifically, several acquired orientation detection images are preprocessed, including image denoising and color equalization. Image denoising is achieved using methods such as Gaussian filtering, while color equalization uses histogram equalization to increase the grayscale differences between pixels, making image details clearer and ensuring a uniform and natural color tone after subsequent stitching and fusion. After preprocessing, feature points are extracted from each preprocessed image using the BRISK algorithm. The BRISK algorithm can accurately capture structural feature points such as the contour and corners of the tool. These feature points provide a basis for the association and matching of different orientation detection images. The extracted feature points are then matched using the FLANN algorithm. The FLANN algorithm, with its feature point matching capability, can quickly filter out corresponding source feature points between different images. Through the correspondence of these source feature points, it accurately determines the spatial transformation relationship between several orientation detection images, covering the relative positions and rotational relationships between images. This clarifies the positional association of the tool structure in each orientation detection image, providing a precise reference for image registration. Based on the determined spatial transformation relationship, orientation detection images with different orientations, positional offsets, or rotational differences are registered using affine transformations, perspective transformations, etc., ensuring that the corresponding tool structures in each image are consistent and avoiding tool structure deformation and information overlap and confusion after merging due to image misalignment. After registration, the pyramid fusion algorithm is used to merge the images. The pyramid fusion algorithm fully preserves the detailed information of the tool in each orientation detection image by decomposing and fusing the images layer by layer, avoiding feature loss caused by the limitation of a single image perspective, while suppressing the risk of image blurring and detail weakening during the fusion process. Finally, the merged orientation detection images are stitched together. The stitching process includes removing stitching gaps and sharpening image edges. Removing stitching gaps eliminates boundary traces at the junctions of different images, making the merged image transition naturally and be coherent overall. Sharpening image edges highlights the contour features of the tool to be detected, enhances the distinction between the tool and the background, and improves the clarity of image details, ensuring the image quality of the image to be detected.

[0033] In some embodiments of this application, when extracting the tool pixels of the image to be detected and performing curve fitting on all tool pixels to determine the pixel position of the tool pixels that are not curve-fitted, the process includes: using a preset deployment time as the X-axis coordinate value of each tool pixel, using the tool pixel value of each tool pixel as the Y-axis coordinate value, establishing a tool coordinate system based on the X-axis and Y-axis coordinate values, converting each tool pixel into a tool pixel coordinate point according to the X-axis and Y-axis coordinate values, performing curve fitting on all tool pixel coordinate points to determine the tool pixel fitting curve, and determining the tool pixel position of the tool pixel coordinate point that is not fitted based on the tool pixel fitting curve.

[0034] Specifically, tool pixels are extracted from the image to be detected. These extracted pixels may still contain redundant pixels such as image noise and background residue. To achieve ordered analysis of the tool pixels, a preset deployment time is used as the X-axis coordinate value for each tool pixel. The preset deployment time is once per second, with one tool pixel extracted every second. The tool pixel value of each tool pixel is used as the Y-axis coordinate value. The preset deployment time assigns a unique temporal identifier to each tool pixel, ensuring that different tool pixels do not repeat in the tool coordinate system. The tool pixel value directly reflects the characteristics of the tool pixel. A tool coordinate system is established based on the set X-axis and Y-axis coordinate values. This tool coordinate system provides a unified analysis benchmark for all tool pixels. By transforming scattered tool pixels into tool pixel coordinates with temporal and feature correlations, the analysis bias caused by the disorder of tool pixels is avoided. Curve fitting is performed using software such as Matlab, or methods such as polynomial fitting and least squares. Since the tool under test has a certain regular shape structure, the corresponding curve fitting results can show a distribution trend that conforms to the tool contour. Based on the tool pixel fitting curve, tool pixel coordinates not covered by the curve are screened out, and the pixel positions of these unfitted tool pixels are determined. Redundant pixels are initially screened to avoid interference with image recognition and to ensure that the subsequent tool detection image only contains tool pixels with real structure, thus improving the accuracy and reliability of subsequent tool parameter analysis.

[0035] In some embodiments of this application, when the image to be detected is substituted into the tool recognition model to determine the predicted pixel position, the process includes: obtaining a tool sample dataset, dividing the tool sample dataset into a training set and a test set, using the training set to train the image network model, using the test set to test the trained image network model, and finally determining a tool recognition model with the image to be detected as the input and the predicted pixel position as the output. The image network model includes a BP neural network model or an RBF neural network model.

[0036] Specifically, the tool sample dataset includes image data under different environments (light, humidity, background interference), pixel positions corresponding to key tool structures (tool tip, cutting edge, tool holder), pixel parameters corresponding to redundant pixels (noise, background residue), and pixel distribution data of tools at different placement angles. The tool sample dataset can comprehensively support the model in learning the distribution patterns of effective and redundant pixels of the tool, ensuring the sufficiency of model training and enabling the model to adapt to the pixel features of different scenarios and different types of tools. The tool sample dataset is divided into training and testing sets. The training set is used for the model to learn feature patterns, while the testing set is used to verify the model's training effect, avoiding the risk of overfitting or underfitting and ensuring the model's generalization ability. The training set is used to train an image network model, either a BP neural network or an RBF neural network. Both models have strong feature learning and fitting capabilities, accurately capturing the positional patterns of tool pixels in the sample data. The training set is then used to test the trained model, optimizing model parameters and correcting prediction biases. Finally, a tool recognition model is determined that takes the image to be detected as input and outputs the predicted pixel positions. Based on the learned patterns, the tool recognition model can accurately output the predicted pixel positions of tool pixels in the image to be detected. The training and testing of the BP or RBF neural network model avoids the bias of subjective human judgment, thereby improving the accuracy and reliability of image recognition of controlled tools.

[0037] In some embodiments of this application, when comparing the pixel position with the predicted pixel position, determining whether to delete the tool pixel based on the comparison result, and determining the tool detection image of the tool to be detected, the process includes: comparing the tool pixel position with the predicted pixel position; if the tool pixel position and the predicted pixel position are consistent, then the corresponding tool pixel is deleted; and the tool detection image of the tool to be detected is determined based on the deletion result; if the tool pixel position and the predicted pixel position are inconsistent, then the image to be detected is determined as the tool detection image of the tool to be detected.

[0038] Specifically, based on the determined tool pixel positions that are not fitted by the curve, and the predicted pixel positions output by the tool recognition model, the tool pixel positions that are not fitted by the curve are candidate targets for suspected redundant pixels, and the predicted pixel positions are the positions of redundant pixels obtained by the model through learning the sample patterns. The two types of locations are compared precisely one by one. If a certain tool pixel position matches the predicted pixel position, it indicates that the tool pixel is a redundant pixel feature, an invalid pixel caused by image noise, background residue, etc., and does not reflect the true structure of the tool to be detected. In this case, the tool pixel is deleted, and the remaining pixel information after deleting redundant pixels is integrated to determine the tool detection image of the tool to be detected. If the tool pixel position does not match the predicted pixel position, it indicates that the tool pixel is a valid pixel of the tool to be detected. The failure to be curve-fitted is due to the local structural details of the tool to be detected, and it is not a redundant pixel. In this case, it does not need to be deleted, and the original image to be detected is determined as the tool detection image of the tool to be detected. By accurately distinguishing between valid pixels and redundant pixels, invalid pixels are avoided from interfering with the subsequent parameter analysis, ensuring that the tool detection image truly and completely reflects the tool structure. At the same time, it avoids the bias of human subjectivity and ensures the accuracy of the identification of controlled tools.

[0039] In some embodiments of this application, when analyzing a tool detection image to determine the tool parameters of the tool to be detected, the process includes: analyzing the tool detection image based on a contour detection algorithm to determine the tool parameters of the tool to be detected, including the tool tip, cutting edge, handle, tail, blood groove, blade, guard, and length.

[0040] Specifically, redundant pixels in the tool detection image have been removed, retaining only the valid pixels of the tool to be detected. A contour detection algorithm is used to extract the contour of the tool detection image, outlining a clear and continuous outline of the tool, forming a contour map containing tool structural information. Based on this contour map, when analyzing the blade tip, the contour detection algorithm tracks the endpoints of the contour lines and selects the endpoint with the largest curvature and a sharp protrusion. This endpoint is the blade tip, characterized by the converging vertex of the contour lines, which can be accurately distinguished from other endpoints by changes in curvature. When analyzing the blade edge, long, strip-shaped lines extending along the blade with continuously changing slopes and sharp edges are identified on the contour. As the cutting part, the blade edge has a smooth contour line and a clear thickness difference from the main body of the blade. When analyzing the handle, the contour detection algorithm locates the area that connects to the blade, has a relatively uniform contour width, and no sharp edges. As the grip part, the handle has a regular contour line and a structural boundary at the connection with the blade. When analyzing the blade tail, the contour detection algorithm traces the extension trajectory of the handle contour to locate the endpoint of the handle furthest from the blade. This endpoint has no sharp features and the contour lines gently end. When analyzing the blood groove, the algorithm identifies the continuous concave line area on the blade contour. The blood groove is a groove structure on the blade, and its contour shows a regular inward concave line, forming a difference in elevation with the main blade contour. When analyzing the blade, the contour area between the blade tip and the guard is defined by the blade tip and guard. The blade is the load-bearing part of the knife, with the largest contour area and connecting the blade tip, blade edge, and guard. When analyzing the guard, the contour detection algorithm identifies the narrow contour area at the junction of the blade and handle. The guard is a structure that separates the blade and handle, and its contour shows a narrow convex or flat transition feature, with a width smaller than that of the blade and handle. When analyzing the blade length, based on the accurately located blade tip and blade tail contours, the contour trajectory between the two points is traced along the blade's central axis. The straight-line distance of the contour trajectory is determined by the contour detection algorithm, which is the blade length. Contour detection algorithms can fully adapt to the features of tool detection images, accurately capture the structural differences of various parameters, and achieve comprehensive and accurate identification of tool parameters. This avoids the shortcoming of only being able to identify the existence of tools without obtaining specific parameters. At the same time, contour detection algorithms avoid the subjective bias of manual measurement, improving the accuracy and reliability of the identification of controlled tools.

[0041] In some embodiments of this application, when comparing the tool parameters with standard tool parameters and determining the corresponding pop-up prompt based on the comparison result, the process includes: comparing the tool parameters with standard tool parameters; if the data in the tool parameters is consistent with the standard data in the standard tool parameters, the tool to be tested is determined to be a controlled tool and a red pop-up prompt is issued; otherwise, the tool to be tested is determined to be a non-controlled tool and a green pop-up prompt is issued.

[0042] Specifically, the standard knife parameters are determined according to the "Standards for the Identification of Controlled Knives" (which covers the rules for determining controlled knives such as daggers, triangular scrapers, and spring knives (switchblades). It covers the control standards corresponding to parameters such as the blade tip, blade edge, and blade length, and is the authoritative basis for determining whether a knife is a controlled knife. The knife parameters are compared with standard knife parameters, covering all dimensions of the knife parameters. If any one or more data points in the knife parameters match the corresponding standard data in the standard knife parameters, it indicates that the structural parameters of the knife under test meet the criteria for a controlled knife. The knife is then identified as a controlled knife, and a red pop-up notification is simultaneously issued, immediately alerting security personnel that the knife is a prohibited item and requires timely interception and disposal. If none of the data points in the knife parameters match the standard data in the standard knife parameters, it indicates that the knife does not meet the criteria for a controlled knife. The knife is then identified as a non-controlled knife, and a green pop-up notification is issued. The green pop-up conveys a safety signal, informing security personnel that the knife can be released normally. Using the "Controlled Knife Identification Standard" instead of subjective human judgment eliminates the risk of bias and misjudgment associated with manual assessment, ensuring the accuracy of the controlled knife identification results. The differentiated notifications of red and green pop-ups allow security personnel to quickly identify the results, improving security efficiency while accurately distinguishing between controlled and non-controlled knives.

[0043] In summary, the beneficial effects of this invention are as follows: In the initial stage of detection, by acquiring environmental detection data of the tool under test based on the acquisition interval, determining environmental index data, and judging the passability of the image detection environment of the tool under test, adverse detection environments are screened in advance, avoiding interference from environmental factors on image acquisition quality. This avoids the risk of insufficient image recognition reliability due to a lack of environmental adaptation mechanisms, providing a stable environmental foundation for image acquisition. Acquiring several orientation detection images and merging them into the image under test integrates tool information from multiple perspectives, compensating for the limitations of a single image's perspective and incomplete features. This allows the image under test to fully present various structural details of the tool under test. After extracting the tool pixels from the image under test, the positions of unfitted pixels are determined through curve fitting. By determining the predicted pixel position using a knife recognition model and removing redundant pixels, invalid information such as noise and background residue in the image can be removed, resulting in a knife detection image that accurately reflects the true structure of the knife to be detected. This solves the risk of feature distortion caused by pixel mixing and improves the accuracy of parameter analysis. By comparing the knife parameters with standard knife parameters, it not only addresses the shortcomings of existing methods in determining the specific parameters of the knife, but also accurately determines whether the knife to be detected is a controlled knife. It replaces the subjectivity and experience dependence of manual judgment, avoiding missed detections and misjudgments. At the same time, red and green pop-up prompts allow security personnel to quickly obtain the judgment results and take corresponding actions, balancing the rigor and smoothness of security inspection work, and providing reliable support for the prevention and control of controlled knives in the field of public safety.

[0044] In another preferred embodiment based on the above embodiments, see [reference] Figure 2-4 As shown, this embodiment provides a smart knife identification device for use with the above-mentioned smart knife identification device for automatic judgment. It includes: a base 1, a display screen 2, a support column 3, a crossbeam 4 and a support block 6. One end of the support column 3 is fixedly connected to the base 1, and the end of the support column 3 away from the base 1 is fixedly connected to the crossbeam 4. One end of the support block 6 is fixedly connected to the support column 3, and the end of the support block 6 away from the support column 3 is fixedly connected to the display screen 2. A camera 5 is installed on the crossbeam 4.

[0045] Specifically, the base 1 serves as the load-bearing foundation, providing stable support for the entire device and adapting to the fixed deployment requirements of security inspection scenarios. The support column 3 acts as a connecting component, with one end fixedly connected to the base 1 and the other end fixedly connected to the crossbeam 4, achieving vertical connection between the upper and lower components. This design is compact and space-efficient. The base 1 and support column 3, support column 3 and crossbeam 4, support column 3 and support block 6, and support block 6 and display screen 2 are all fixedly connected, avoiding the risk of component loosening or displacement during high-frequency use in security inspection scenarios. This ensures the structural stability of the equipment during long-term operation and reduces the risk of image acquisition and detection errors caused by structural displacement. The display screen 2 is positioned in an easily observable area for operators, and its angled design allows security personnel to view the detection results in real time (red / green pop-up prompts). The crossbeam 4 is located at the top of the support column 3, and the camera 5 is integrated inside the crossbeam 4. This allows the camera 5 to obtain an appropriate shooting height and field of view, thereby efficiently acquiring several directional detection images of the tool to be inspected, providing hardware support for automatic image acquisition and processing.

[0046] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do 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. An automatic determination method for a smart knife identification device, characterized in that, include: The tool to be inspected is identified, and several environmental detection data of the tool to be inspected are acquired based on the acquisition interval. Environmental index data are determined based on environmental detection data of the same type. The passability of the image detection environment of the tool to be inspected is judged based on all environmental index data. If the image detection environment of the tool to be detected is qualified, then several orientation detection images of the tool to be detected are determined, and the several orientation detection images are merged to determine the image to be detected of the tool to be detected. Extract the knife pixel points of the image to be detected, and perform curve fitting on all knife pixel points to determine the pixel positions of the knife pixel points that are not curve fitted. Substitute the image to be detected into the knife recognition model to determine the predicted pixel positions, and compare the pixel positions with the predicted pixel positions. Based on the comparison results, determine whether to delete the knife pixel points to determine the knife detection image of the knife to be detected. The tool detection image is analyzed to determine the tool parameters of the tool to be detected, and the tool parameters are compared with the standard tool parameters. Based on the comparison results, a corresponding pop-up prompt is generated.

2. The automatic determination method of the intelligent identification device for controlled knives according to claim 1, characterized in that, When acquiring several environmental detection data points of the tool under test based on the acquisition interval, and determining environmental indicator data based on environmental detection data of the same type, the process includes: Obtain all environmental monitoring data of the same type, and determine the average value of all environmental monitoring data of the same type as the environmental indicator data.

3. The automatic determination method of the intelligent identification device for controlled knives according to claim 2, characterized in that, When determining the passability of the image detection environment of the tool under test based on all environmental index data, the following are included: Determine the range of environmental indicator data corresponding to each environmental indicator data, and compare each environmental indicator data with its corresponding range of environmental indicator data; If each environmental indicator data is within the corresponding environmental indicator data range, then the image detection environment of the tool to be tested is deemed qualified. If one or more environmental indicator data are within the corresponding environmental indicator data range, then the image detection environment of the tool to be tested is determined to be unqualified.

4. The automatic determination method of the intelligent identification device for controlled knives according to claim 3, characterized in that, When merging several orientation detection images to determine the image to be detected for the cutting tool, the process includes: Several orientation detection images are preprocessed, including image denoising and color equalization. Feature points are extracted from the preprocessed orientation detection images using the BRISK algorithm, and the extracted feature points are matched using the FLANN algorithm to determine the spatial transformation relationship between the preprocessed orientation detection images. The spatial transformation relationship includes relative position and rotation relationship. The preprocessed orientation detection images are registered based on the spatial transformation relationship, and the registered orientation detection images are merged using the pyramid fusion algorithm. The merged orientation detection images are then stitched together to determine the image to be detected for the tool to be detected. The stitching process includes removing stitching gaps and sharpening image edges.

5. The automatic determination method of the intelligent identification device for controlled knives according to claim 4, characterized in that, When extracting the tool pixels from the image to be detected, and performing curve fitting on all tool pixels to determine the pixel positions of tool pixels that are not curve-fitted, the process includes: The preset deployment time is used as the X-axis coordinate value of each tool pixel, and the tool pixel value of each tool pixel is used as the Y-axis coordinate value. A tool coordinate system is established based on the X-axis coordinate value and the Y-axis coordinate value, and each tool pixel is converted into a tool pixel coordinate point according to the X-axis coordinate value and the Y-axis coordinate value. Curve fitting is performed on all tool pixel coordinate points to determine the tool pixel fitting curve, and the tool pixel position of the unfitted tool pixel coordinate points is determined based on the tool pixel fitting curve.

6. The automatic determination method of the intelligent identification device for controlled knives according to claim 5, characterized in that, When substituting the image to be detected into the tool recognition model to determine the predicted pixel position, the following steps are included: Obtain a tool sample dataset and divide the tool sample dataset into a training set and a test set; The image network model is trained using the training set and tested using the test set. Finally, the input is determined to be the image to be detected, and the output is the tool recognition model with the predicted pixel position. The image network model includes a BP neural network model or an RBF neural network model.

7. The automatic determination method of the intelligent identification device for controlled knives according to claim 6, characterized in that, When comparing the stated pixel position with the predicted pixel position, and determining whether to delete the tool pixel based on the comparison result to determine the tool detection image of the tool to be detected, the process includes: The tool pixel position is compared with the predicted pixel position; If the tool pixel position is consistent with the predicted pixel position, the corresponding tool pixel is deleted, and the tool detection image of the tool to be detected is determined based on the deletion result. If the tool pixel position and the predicted pixel position are inconsistent, the image to be detected is determined as the tool detection image of the tool to be detected.

8. The automatic determination method of the intelligent identification device for controlled knives according to claim 7, characterized in that, When analyzing the tool detection image to determine the tool parameters of the tool to be detected, the following steps are included: The tool detection image is analyzed based on the contour detection algorithm to determine the tool parameters of the tool to be detected. The tool parameters include the tip, cutting edge, handle, tail, blood groove, blade, guard, and length.

9. The automatic determination method of the intelligent identification device for controlled knives according to claim 8, characterized in that, When comparing the tool parameters with standard tool parameters and determining the corresponding pop-up prompt based on the comparison result, the following are included: The tool parameters are compared with the standard tool parameters; If the data in the tool parameters is consistent with the standard data in the standard tool parameters, the tool to be tested is determined to be a controlled tool and a red pop-up window is issued; otherwise, the tool to be tested is determined to be a non-controlled tool and a green pop-up window is issued.

10. A smart knife identification device, used for applying the automatic determination method of the smart knife identification device as described in any one of claims 1-9, characterized in that, include: Base, display screen, support columns, crossbeams, and support blocks; One end of the support column is fixedly connected to the base, and the other end of the support column away from the base is fixedly connected to the crossbeam; One end of the support block is fixedly connected to the support column, and the other end of the support block away from the support column is fixedly connected to the display screen. A camera is installed on the crossbeam.

Images