Camera quality detection method and system based on data analysis

Abstract

The invention discloses a camera quality detection method and system based on data analysis, and relates to the technical field of camera quality detection, and the system comprises a position determination module which is used for determining the positions of a camera and a to-be-shot image; the imaging feedback module is used for acquiring a pattern image of the camera and feeding the pattern image back to the system platform to splice a Sudoku image; the pattern recognition module is used for recognizing the sudoku image based on a sudoku image algorithm and matching a pattern recognition result with pattern information stored offline in advance; and the quality detection module is used for mapping the condition of abnormal pattern imaging to a corresponding camera index directory, determining the number of the camera for shooting, and feeding back the quality problem of the corresponding camera. According to the method, the group of patterns are shot through each group of cameras, the structure information of each pattern picture is recognized through an algorithm, the pattern image with problems is obtained, and the subdivision quality problem occurring in the cameras is recognized.

Description

Technical Field

[0001] The present invention relates to the technical field of camera quality detection, and in particular to a camera quality detection method and system based on data analysis. Background Art

[0002] Some quality problems may occur in the camera production process. Common camera quality problems will cause abnormal imaging images, which are specifically manifested as: the image becomes black and white mode, the image is mirrored, the color is cast or distorted, data is lost, etc.

[0003] Current camera quality testing technology still has many shortcomings. For example, camera testing generally uses feature extraction for detection and analysis. For example, Patent Publication No. CN119854481A, "Camera Testing Method, Apparatus, Device, and Program Product," states that by acquiring a target camera's target surveillance image and text prompts, features are extracted from the target surveillance image and text prompts. These features are then input into a target deep learning model to predict image quality. The target camera is then tested based on the predicted image quality. However, this method has several shortcomings. For example, the text prompts are limited in their ability to describe image quality metrics. If the text prompts are not comprehensive or accurate in their description of image quality, this can lead to biased assessments of camera image quality. For example, some complex image quality issues may not be accurately expressed using the pre-set text prompts, thus affecting the accuracy of the test results. Furthermore, limited sample data can prevent accurate identification of special scenarios or rare camera failures. For example, when testing image quality under extreme lighting conditions, the model may not be able to provide an accurate assessment if the sample lacks relevant data. Furthermore, the number and consistency of modules are key considerations during the inspection process. For example, in the invention patent "Camera Module Inspection Method, Apparatus, and Device," published under the publication number "CN119835406A," a standard camera module is determined based on a set of multispectral images captured by different numbers of camera modules. The standard multispectral image of the standard camera module is then compared for consistency with each image in the other set of multispectral images to determine whether the module is defective. This approach addresses the issue of low camera module inspection efficiency. The selection of the standard module is affected by the number of samples. If the number of camera modules in the first set is small, it may not accurately represent the overall module performance, leading to biased standard module selection. For example, if the first set of modules contains some modules with good or poor performance, this can affect the accuracy of the standard modules, further impacting the inspection results of other modules. Secondly, regarding the module consistency assumption, this method assumes that the first set of camera modules is ideally highly consistent. However, in actual production, even camera modules from the same batch can exhibit certain individual differences. This assumption may lead to ignoring these differences when determining the standard module, making subsequent consistency comparisons less accurate. Summary of the Invention

[0004] The purpose of the present invention is to provide a camera quality detection method and system based on data analysis to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a camera quality detection method based on data analysis, the method comprising: Determine the position of the camera and the image to be captured; Obtain the pattern imaging of the camera and feed it back to the system platform to stitch it into a nine-square grid image; The nine-square grid image is recognized based on the nine-square grid image algorithm, and the pattern recognition result is matched with the pattern information stored offline in advance. If a match is found, it means that the pattern imaging is normal, and if a match is not found, it means that the pattern imaging is abnormal; Combine the loss functions of each task by weight, ,in, 、 、 Refers to the loss function , loss function , loss function The weight coefficients of , add up to 1; by balancing the weights, the training gradients are different from any other; After the backbone network outputs the candidate boxes, Soft-NMS is used to process overlapping detection boxes and retain boxes with high confidence and better shapes; Color block filtering calculates the color difference between the average Lab value of the detected color block and the standard value. If the color difference is greater than the system-set threshold, the current detection result is filtered. The standard value is obtained by extracting the RGB value of each color block in the same category of color block areas marked in the collected positive sample data sets. The RGB value of a single color block is converted to Lab space representation through a color space conversion model. The Lab components of all similar color blocks are averaged, and the final Lab average value is used as the standard value of the color block of that category. Character filtering: In mirrored scenes, the character outline is judged by its left-right symmetry. If it is a mirror image, the character detection result is filtered; Based on the overall target and part target information of each grid image in the nine-grid image, the overall target and part target recognition category information are classified according to the spatial position relationship between the overall target frame and the part target frame; According to the part target frame and the overall target frame, calculate the ratio of the intersection area of ​​the part and the overall target to the part frame; ; in, Indicates the ratio of the intersection area of ​​the part and the overall target to the part box; Indicates the Coordinate data of each part; Indicates the The coordinate data of the overall target frame; Indicates the overlapping area of ​​the part frame and the overall target frame; Indicates the The width of the frame of each part; Indicates the height of the i-th part box, 、 Represent the quantity value of the part and the quantity value of the overall target frame respectively. If , where t is the preset threshold, , then the i-th part is determined to belong to the category corresponding to the j-th overall target; otherwise, it is determined not to belong.

[0006] According to the above technical solution, candidate parts and overall targets are classified by the ratio of the intersection area of ​​the part and the overall target to the part frame. The candidate part frames classified into this category are sorted and output in the spatial order of upper left corner, upper right corner, lower right corner, and lower left corner. The part category name is added with the category prefix and output as a character string together with the overall target category.

[0007] According to the above technical solution, the system determines the color of the pattern and the character string in each grid respectively, outputs the algorithm recognition result string based on the nine-grid image algorithm recognition model, recorded as the first string, obtains the string of the system calibration pattern information, recorded as the second string, calculates the editing distance required to convert the first string to the second string, and obtains the length of the first string and the second string at the same time, and calculates the final matching degree.

[0008] According to the above technical solution, the calculation of the edit distance includes deletion and insertion operations. In the process of converting the first string to be exactly the same as the second string, the deletion and insertion operations at the same position are recorded as one distance.

[0009] According to the above technical solution, the calculation of the final matching degree includes: ; in, Refers to the final matching degree; Refers to the length of the first string and the second string respectively; Refers to the edit distance.

[0010] A camera quality detection system based on data analysis, the system comprising: A position determination module, used to determine the position of the camera and the image to be captured; The imaging feedback module is used to obtain the pattern imaging of the camera and feed it back to the system platform to stitch it into a nine-square grid image; The pattern recognition module recognizes the nine-square grid image based on the nine-square grid image algorithm and matches the pattern recognition results with the pattern information stored offline in advance. If a match is found, it indicates that the pattern imaging is normal; if a match is not found, it indicates that the pattern imaging is abnormal; The quality detection module is used to map abnormal pattern imaging to the corresponding camera index directory, determine the camera number of the shooting, and feedback the quality problem of the corresponding camera.

[0011] Compared with the existing technology, the beneficial effects of the present invention are: the present invention constructs a set of standard patterns and calibrates the pattern information offline in advance, shoots the set of patterns through each group of cameras, identifies the structural information of each pattern picture through an algorithm, and matches the structural information with the above-mentioned calibrated pattern information to obtain a pattern image with problems, indicating that there is a problem with the corresponding camera for shooting, and can identify the subdivision quality problems of the camera. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A schematic diagram of the steps of a camera quality detection method based on data analysis of the present invention; Figure 2 This is a schematic diagram of the setting of one of the nine-square grids of a camera quality detection method based on data analysis of the present invention. DETAILED DESCRIPTION

[0013] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0014] Example: Figure 1-Figure 2 As shown, the present invention provides a camera quality detection method based on data analysis, the method comprising: Determine the position of the camera and the image to be captured; Obtain the pattern imaging of the camera and feed it back to the system platform to stitch it into a nine-square grid image; The nine-square grid image is recognized based on the nine-square grid image algorithm, and the pattern recognition result is matched with the pattern information stored offline in advance. If a match is found, it means that the pattern imaging is normal, and if a match is not found, it means that the pattern imaging is abnormal; Map abnormal pattern imaging to the corresponding camera index directory, determine the camera number of the shooting, and feedback the quality problem of the corresponding camera.

[0015] Determining the position of the camera and the image to be captured includes: Take any set of patterns and paste them from left to right on the same plane at the same height according to the pattern serial number, and fix the patterns at a fixed interval; place a group of cameras to be detected at the same height as the center of the pattern at a fixed distance in front of the pattern. Each camera to be detected corresponds to one pattern. By setting the distance between the camera to be detected and the pattern, as well as the fixed interval between the patterns, it is ensured that any camera to be detected can only capture the corresponding pattern.

[0016] The method of obtaining the pattern imaging of the camera and feeding it back to the system platform to form a nine-square grid image includes: Set the label category for the overall pattern target in the nine-square grid. The label categories of the rectangular boxes from 1 to 9 are defined as pattern_1 to pattern_9 respectively; the image features in any square must be labeled with the corresponding square label category.

[0017] Also includes: Based on the determined camera and image position, sample data is obtained, model training is performed based on the sample data, and a nine-square image algorithm recognition model is generated, specifically including: The annotation scope is limited, and several positive sample data sets are collected. A single positive sample data is a nine-square grid image of a pattern image captured by a group of cameras. The annotation strategy for each square image in the nine-square grid is to only define the target and part. The annotation scope is limited as follows: The overall goal is to mark the minimum enclosing rectangle of the complete pattern; The part target includes the color blocks at the four corners of the grid and the characters in the middle. The color blocks at the four corners of the grid include rectangles marked with fixed colors, recording the coordinates and RGB values ​​and categories of all pixels in the area, where the categories are: blue, yellow, red, and green; the characters in the middle include: the minimum circumscribed rectangle marked with numbers 1 to 9, recording the character category, which are: 1, 2, 3...9, and its contour key points.

[0018] The training process includes: A data augmentation strategy is used to achieve simulated mirror imaging: the positive sample image is horizontally flipped, retaining the overall and color block annotations; Simulate color cast: Randomly shift the HSV color space of the positive sample image to weaken the model's dependence on fixed colors; Simulating partial incompleteness: Randomly occlude 10%-40% of the part area in the positive sample image to learn part integrity judgment; Simulated grayscale conversion: converting positive sample images to grayscale, retaining character and overall shape annotations, with the goal of enhancing shape feature learning; Design loss function: In terms of the overall goal, Focal Loss+CIoU Loss is used to construct the loss function , strengthen bounding box regression and category balance issues; In terms of color blocks, CIoU Loss+L2 Loss is used to construct the loss function , CIoU Loss is used to calculate the regression loss of the color block detection box. The color loss is calculated by converting the enhanced image RGB space and the original image RGB space into Lab space. The color loss of the color cast sample is automatically increased to suppress false detection; In terms of characters, CIoU Loss+Smooth L1 Loss is used to construct the loss function , use CIoULoss to calculate the regression loss of the character detection box, and use Smooth L1 Loss to implement the stroke inflection point coordinate regression loss; Combine the loss functions of each task by weight, , by balancing the different training gradients by weights; After the backbone network outputs the candidate boxes, Soft-NMS is used to process overlapping detection boxes and retain boxes with high confidence and better shapes; Color block filtering calculates the color difference between the average Lab value of the detected color block and the standard value. If the color difference is greater than the system-set threshold, the current detection result is filtered. The standard value is obtained by extracting the RGB value of each color block in the same category of color block areas marked in the collected positive sample data sets. The RGB value of a single color block is converted to Lab space representation through a color space conversion model. The Lab components of all similar color blocks are averaged, and the final Lab average value is used as the standard value of the color block of that category. Character filtering: In mirrored scenes, the character outline is judged by its left-right symmetry. If it is a mirror image, the character detection result is filtered; Based on the overall target and part target information of each grid image in the nine-grid image, the overall target and part target recognition category information are classified according to the spatial position relationship between the overall target frame and the part target frame; According to the part target frame and the overall target frame, calculate the ratio of the intersection area of ​​the part and the overall target to the part frame; ; in, Indicates the ratio of the intersection area of ​​the part and the overall target to the part box; Indicates the Coordinate data of each part; Indicates the The coordinate data of the overall target frame; Indicates the overlapping area of ​​the part frame and the overall target frame; Indicates the The width of the frame of each part; Indicates the height of the i-th part box. 、 Represent the quantity value of the part and the quantity value of the overall target frame respectively. If , where t is the preset threshold, , the present invention sets the corresponding value to 0.8, and determines that the i-th part belongs to the category corresponding to the j-th overall target; otherwise, it is determined that it does not belong.

[0019] The candidate parts and the overall target are classified by the ratio of the intersection area of ​​the part and the overall target to the part frame. The candidate part frames classified into this category are sorted and output in the spatial order of upper left corner, upper right corner, lower right corner, and lower left corner. The part category name is added with the category prefix and output as a character string together with the overall target category.

[0020] The system determines the color of the pattern and the character string in each grid respectively, and outputs the algorithm recognition result string based on the nine-grid image algorithm recognition model, which is recorded as the first string. The string of the system-calibrated pattern information is obtained, which is recorded as the second string. The editing distance required to convert the first string to the second string is calculated, and the lengths of the first string and the second string are obtained at the same time to calculate the final matching degree.

[0021] The calculation of the edit distance includes deletion and insertion operations. In the process of converting the first character string to be identical to the second character string, deletion and insertion operations at the same position are recorded as one distance.

[0022] The calculation of the final matching degree includes:

[0023] in, Refers to the final matching degree; Refers to the length of the first string and the second string respectively; Refers to the edit distance.

[0024] In this embodiment, a camera quality detection system based on data analysis is also included, and the system includes: A position determination module, used to determine the position of the camera and the image to be captured; The imaging feedback module is used to obtain the pattern imaging of the camera and feed it back to the system platform to stitch it into a nine-square grid image; The pattern recognition module recognizes the nine-square grid image based on the nine-square grid image algorithm and matches the pattern recognition results with the pattern information stored offline in advance. If a match is found, it indicates that the pattern imaging is normal; if a match is not found, it indicates that the pattern imaging is abnormal; The quality detection module is used to map abnormal pattern imaging to the corresponding camera index directory, determine the camera number of the shooting, and feedback the quality problem of the corresponding camera.

[0025] In a specific embodiment of the present application, the pattern is designed as a circular cake with a number in the center. The circular cake is divided into four equal parts with different color areas in the center. The four corners of the whole pattern have four different color areas of equal size. The four corner color blocks are connected to the colored circular cake. The present invention designs a group of patterns with the center numbers 1-9 respectively and the other areas are completely the same. For example, the schematic diagram of pattern No. 1 is as follows Figure 2 As shown; The numbered design in the center facilitates the mapping between patterns and cameras. For example, a group of cameras is mounted in fixed positions, each numbered 1 through 9. The corresponding pattern images captured by each camera are stitched together into a nine-square grid, with images of patterns 1 through 9 placed in squares 1 through 9, respectively. Therefore, the pattern position numbers correspond one-to-one with the camera mount position numbers. This allows for quick identification of camera quality issues when communicated via voice or visual means. The numbered design also reduces the difficulty of algorithmic recognition. The four corners are designed with rectangular blocks of different colors to prevent image distortion caused by black and white mode, mirroring, color cast or distortion, or data loss, which can affect recognition results in these areas and identify camera quality issues.

[0026] At the same time, the present invention prints out the group of patterns using A3 paper, and the patterns No. 1-9 are pasted on the wall at the same height and from left to right, with the intervals between the patterns fixed; a group of cameras with the same height as the center of the pattern are placed at a certain distance in front of the pattern, and each camera corresponds to one pattern. By setting the distance between the camera and the pattern, as well as the distance between the patterns, it is ultimately ensured that each camera only captures the corresponding pattern image. Based on the above operations, the pattern position, interval, and camera placement position are also fixed. In this way, when each group of cameras of the same model is detected, the camera is placed in a fixed position for imaging.

[0027] Based on the above operations, when each camera captures the corresponding pattern, there is only one pattern in the image, and the pattern occupies the majority of the image area. In addition, the background where the pattern is pasted should be as pure as possible, with a different background color from the pattern color, and uniform lighting should be guaranteed.

[0028] Based on the aforementioned camera and pattern deployment operations, a large number of nine-grid images are collected, and then those with acceptable imaging quality are manually selected. Based on the aforementioned selection algorithm, the samples need to be labeled for overall targets and parts. At the same time, in order to automatically distinguish the pattern information of each grid image, the overall target of each pattern can be designed into different label categories, and the recognition results can be automatically divided by label category. The labeling rules are detailed as follows: Each pattern in the nine-grid image needs to be distinguished, so the label categories for each pattern are set differently. The label categories for the rectangular boxes in grids 1 to 9 are pattern_1 to pattern_9, respectively. The pattern has five distinct parts, and the label categories for the rectangular boxes in each part are also set differently. For example, in pattern 1, the label categories for the top-left, top-right, bottom-right, and bottom-left color blocks are pattern_1_blue, pattern_1_yellow, pattern_1_red, and pattern_1_green, respectively. The character block in the center of the pattern is labeled pattern_1_1, and the label information for the other patterns is similar.

[0029] The nine-square grid image spliced ​​by the camera group to be detected is transmitted and input into the above model to obtain the overall pattern target information of each square image in the nine-square grid image, as well as the part target information. The target information includes rectangular frame space information and label information.

[0030] By predicting the label information of the overall pattern target of each grid in the nine-grid image, the prediction results can be automatically classified. For example, they can be classified as pattern_1, pattern_2, ..., pattern_9.

[0031] At the same time, the predicted part information is classified according to the same prefix (pattern_n). For example, pattern_1_blue, pattern_1_yellow, pattern_1_red, pattern_1_green, and pattern_1_1 have the prefix pattern_1, which are classified into one category, and are pattern_1 categories. At the same time, the predicted part categories are sorted according to the mutual relationships of the upper left corner, upper right corner, lower right corner, lower left corner, and the middle. The pattern information that has been calibrated offline in advance mainly includes the color block category information of the four corners of the pattern, and the character information in the middle of the pattern. The order of the calibrated color blocks is: upper left color block, upper right color block, lower right color block, and lower left color block. The calibrated pattern information is: upper left color block category, upper right color block category, lower right color block category, lower left color block category, and middle character information. Figure 2 For example, the calibrated pattern information is: pattern_1 / pattern_1_blue / pattern_1_yellow / pattern_1_red / pattern_1_green / pattern_1_1 Perform string matching between the algorithm recognition results and the pattern information previously calibrated offline, and calculate the matching degree: First, the edit distance is calculated using the edit distance algorithm; the matching degree is calculated; finally, the backtracking editing operation path of the edit distance is calculated, all replacement, insertion or deletion positions are marked, and the unmatched character positions and characters are found.

[0032] Specifically, Figure 2 For example, Algorithm recognition result information (first string): pattern_1 / pattern_1_red / pattern_1_green / pattern_1_blue / pattern_1_yellow / pattern_1_1 Calibration pattern information (second string): pattern_1 / pattern_1_blue / pattern_1_yellow / pattern_1_red / pattern_1_green / pattern_1_1 Calculate the edit distance: The edit distance between the first string and the second string is 4.

[0033] Four operations are required, changing pattern_1 / pattern_1_blue to pattern_1 / pattern_1_red, and so on.

[0034] Calculate the matching degree:

[0035] The mismatched locations include: Index 1: pattern_1_blue -> pattern_1_red Index 2: pattern_1_yellow -> pattern_1_green Index 3: pattern_1_red -> pattern_1_blue Index 4: pattern_1_green-> pattern_1_yellow By calculating the matching degree based on the above method, we can determine whether the pattern image captured by the camera is normal. By calculating and finding the mismatching position strings, we can determine which sub-problems exist in the camera quality problem.

[0036] The specific manifestations are: black and white picture, color distortion or color cast, picture mirroring, and data loss.

[0037] The camera is analyzed based on the matching degree calculated above, and the recognition results are matched with the pattern calibration information. If the following mismatch occurs, it indicates a problem with the camera: Therefore, based on the calculated matching degree of the present application, a set of initial test thresholds is constructed. The method for constructing the initial test thresholds includes: Obtaining data on a proportional relationship between the edit distance and the maximum distance when a serious problem occurs with the camera under historical data, that is, when the pattern image captured by the camera is a black and white image or the image has serious color distortion or serious color cast, to form a first data list; Based on the first data list, a set of grey cumulative models is formed, namely, A new set of data lists is formed, in which Represents the nth group of data in the new data list; i represents the sequence number; Represents the i-th data in the first data list; The new data list is processed by weighted neighbor value, and a whitening differential equation is established for the new data list, and finally the model output value based on the first data list is formed: ; in, represents the model output value based on the first data list; represents the parameter vector to be estimated, which is the ratio of the endogenous control gray number to the development gray number in the whitened differential equation; Refers to the i data in the first data list; is the base of the natural logarithm function; N refers to the number of data in the first data list.

[0038] The output As the matching threshold, that is, when the matching When the matching degree threshold is greater than 0, a camera with serious quality problems is predicted.

[0039] That is, when the matching threshold Match , it predicts cameras with fewer quality defects, including cameras with slight distortion, slight color cast, or slight data loss.

[0040] Taking the data of this embodiment as an example, As the matching threshold, if a camera matches , that is, the matching degree is equal to the matching degree threshold, and the recognition result does not have a color category string, that is, the recognized color block information does not match the calibrated color block information at all; but the recognition result has a character category string, and it can match the calibrated character category string, then it is predicted that the pattern image captured by the camera is a black and white image or the image has severe color distortion or severe color cast; If a camera matches 0, and the recognition result contains a color category string but no character category string, then the image of the pattern captured by the camera is predicted to be a mirror image; If a camera matches 0, and there is no color category string in the recognition result, and no character category string in the recognition result, then it is predicted that there will be serious data loss in the pattern image captured by the camera.

[0041] When the edit distance is 1, in this embodiment, if a camera matches , and the recognition result contains color category strings and character category strings, and one of the target prediction category strings in the entire part information does not match the calibration information string, then the camera is predicted to have slight distortion, slight color cast, or slight data loss.

[0042] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A camera quality detection method based on data analysis, characterized by: The method includes: Determine the position of the camera and the image to be captured; Obtain the pattern imaging of the camera and feed it back to the system platform to stitch it into a nine-square grid image; The nine-square grid image is recognized based on the nine-square grid image algorithm, and the pattern recognition result is matched with the pattern information stored offline in advance. If a match is found, it means that the pattern imaging is normal, and if a match is not found, it means that the pattern imaging is abnormal; Map abnormal pattern imaging to the corresponding camera index directory, determine the camera number of the shooting, and feedback the quality problem of the corresponding camera.

2. The camera quality detection method based on data analysis according to claim 1, characterized in that: Determining the position of the camera and the image to be captured includes: Take any set of patterns and paste them from left to right on the same plane at the same height according to the pattern serial number, and fix the patterns at a fixed interval; place a group of cameras to be detected at the same height as the center of the pattern at a fixed distance in front of the pattern. Each camera to be detected corresponds to one pattern. By setting the distance between the camera to be detected and the pattern, as well as the fixed interval between the patterns, it is ensured that any camera to be detected can only capture the corresponding pattern.

3. The camera quality detection method based on data analysis according to claim 1, characterized in that: The method of obtaining the pattern imaging of the camera and feeding it back to the system platform to form a nine-square grid image includes: Set the label category for the overall pattern target in the nine-square grid. The label categories of the rectangular boxes from 1 to 9 are defined as pattern_1 to pattern_9 respectively; the image features in any square must be labeled with the corresponding square label category.

4. The camera quality detection method based on data analysis according to claim 1, characterized in that: Also includes: Based on the determined camera and image position, sample data is obtained, model training is performed based on the sample data, and a nine-square image algorithm recognition model is generated, specifically including: The annotation scope is limited, and several positive sample data sets are collected. A single positive sample data is a nine-square grid image of a pattern image captured by a group of cameras. The annotation strategy for each square image in the nine-square grid is to only define the target and part. The annotation scope is limited as follows: The overall goal is to mark the minimum enclosing rectangle of the complete pattern; The part target includes the color blocks at the four corners of the grid and the middle character. The color blocks at the four corners of the grid include rectangles marked with fixed colors, recording the coordinates and RGB values ​​and categories of all pixels in the area, where the categories are: blue, yellow, red, and green; the middle character includes: the minimum circumscribed rectangle marked with the numbers 1-9, recording the character category, which are: 1, 2, 3...9, and its contour key points.

5. The camera quality detection method based on data analysis according to claim 4, characterized in that: The training process includes: A data augmentation strategy is used to achieve simulated mirror imaging: the positive sample image is horizontally flipped, retaining the overall and color block annotations; Simulate color cast: Randomly shift the HSV color space of the positive sample image to weaken the model's dependence on fixed colors; Simulating partial incompleteness: Randomly occlude 10%-40% of the part area in the positive sample image to learn part integrity judgment; Simulated grayscale conversion: converting positive sample images to grayscale, retaining character and overall shape annotations, with the goal of enhancing shape feature learning; Design loss function L: In terms of the overall goal, Focal Loss+CIoU Loss is used to construct the loss function , strengthen bounding box regression and category balance issues; In terms of color blocks, CIoU Loss+L2 Loss is used to construct the loss function , CIoU Loss is used to calculate the regression loss of the color block detection box. The color loss is calculated by converting the enhanced image RGB space and the original image RGB space into Lab space. The color loss of the color cast sample is automatically increased to suppress false detection; In terms of characters, CIoU Loss+Smooth L1 Loss is used to construct the loss function , use CIoU Loss to calculate the regression loss of the character detection box, and use Smooth L1 Loss to implement the stroke inflection point coordinate regression loss; Combine the loss functions of each task according to the weights, ,in, 、 、 Refers to the loss function , loss function , loss function The weight coefficients of , add up to 1; by balancing the weights, the training gradients are different from any other; After the backbone network outputs the candidate boxes, Soft-NMS is used to process overlapping detection boxes and retain high confidence boxes; Color block filtering calculates the color difference between the average Lab value of the detected color block and the standard value. If the color difference is greater than the system-set threshold, the current detection result is filtered. The standard value is obtained by extracting the RGB value of each color block in the same category of color block areas marked in the collected positive sample data sets. The RGB value of a single color block is converted to Lab space representation through a color space conversion model. The Lab components of all similar color blocks are averaged, and the final Lab average value is used as the standard value of the color block of that category. Character filtering: In mirrored scenes, the character outline is judged by its left-right symmetry. If it is a mirror image, the character detection result is filtered; Based on the overall target and part target information of each grid image in the nine-grid image, the overall target and part target recognition category information are classified according to the spatial position relationship between the overall target frame and the part target frame; According to the part target frame and the overall target frame, calculate the ratio of the intersection area of ​​the part and the overall target to the part frame; ； in, Indicates the ratio of the intersection area of ​​the part and the overall target to the part box; Indicates the Coordinate data of each part; Indicates the The coordinate data of the overall target frame; Indicates the overlapping area of ​​the part frame and the overall target frame; Indicates the The width of the frame of each part; Indicates the height of the i-th part box; 、 Represent the quantity value of the part and the quantity value of the overall target frame respectively. If , where t is the preset threshold, , then the i-th part is determined to belong to the category corresponding to the j-th overall target; otherwise, it is determined not to belong.

6. The camera quality detection method based on data analysis according to claim 5, characterized in that: The candidate parts and the overall target are classified by the ratio of the intersection area of ​​the part and the overall target to the part frame. The candidate part frames classified into this category are sorted and output in the spatial order of upper left corner, upper right corner, lower right corner, and lower left corner. The part category name is added with the category prefix and output as a character string together with the overall target category.

7. The camera quality detection method based on data analysis according to claim 1, characterized in that: The system determines the color of the pattern and the character string in each grid respectively, and outputs the algorithm recognition result string based on the nine-grid image algorithm recognition model, which is recorded as the first string. The string of the system-calibrated pattern information is obtained, which is recorded as the second string. The editing distance required to convert the first string to the second string is calculated, and the lengths of the first string and the second string are obtained at the same time to calculate the final matching degree.

8. The camera quality detection method based on data analysis according to claim 1, characterized in that: The calculation of the edit distance includes deletion and insertion operations. In the process of converting the first character string to be identical to the second character string, deletion and insertion operations at the same position are recorded as one distance.

9. The camera quality detection method based on data analysis according to claim 1, characterized in that: The calculation of the final matching degree includes: ； in, Refers to the final matching degree; Refers to the length of the first string and the second string respectively; Refers to the edit distance.

10. A camera quality detection system based on data analysis, characterized by: The system comprises: A position determination module, used to determine the position of the camera and the image to be captured; The imaging feedback module is used to obtain the pattern imaging of the camera and feed it back to the system platform to stitch it into a nine-square grid image; The pattern recognition module recognizes the nine-square grid image based on the nine-square grid image algorithm and matches the pattern recognition results with the pattern information stored offline in advance. If a match is found, it indicates that the pattern imaging is normal; if a match is not found, it indicates that the pattern imaging is abnormal; The quality detection module is used to map abnormal pattern imaging to the corresponding camera index directory, determine the camera number of the shooting, and feedback the quality problem of the corresponding camera.

Images