A data processing system based on plastic product detection

Abstract

The present invention relates to the field of plastic product detection technology, specifically a data processing system based on plastic product detection, including an image processing unit, a curved surface detection unit, a defect mapping unit and a plane detection unit, and also includes: an image acquisition unit, the image acquisition unit is used to acquire images of target plastic products based on an industrial camera with a fixed position to obtain a target plastic product image, and transmit the target plastic product image to a target detection unit. The present invention detects the acquired target plastic product image based on a trained target detection model by a target detection unit, and can accurately separate the curved surface plastic product sub-image and the flat plastic product sub-image therein, which helps to use different detection methods for different types of plastic products in the future, avoids the inaccuracy that may be caused by unified detection, and makes the detection more targeted.

Description

Technical Field

[0001] The present invention relates to the technical field of plastic product detection, and in particular to a data processing system based on plastic product detection. Background Art

[0002] Plastic products include plastic products with curved shapes, and during the production process, these plastic products may have various quality problems. For example, curved plastic products may have problems such as inaccurate curved contours and uneven surfaces, and flat plastic products may have defects such as scratches, holes, uneven colors or inconsistent thickness.

[0003] Traditional systems often use a unified inspection method for curved and flat plastic products, without distinguishing between the differences in shape, structure and defect type between the two. For example, when inspecting curved plastic products, inspection technology suitable for flat surfaces may be used, which will lead to inaccurate detection of special properties such as the contour and curvature of the curved surface. Similarly, for flat plastic products, since there is no optimized inspection method specifically for flat characteristics, some subtle defects such as shallow scratches or tiny color differences may not be accurately detected; and traditional systems may only perform inspections based on some simple image features, such as overall grayscale values, simple shape template matching, etc. This simple feature analysis is difficult to fully capture the complex features of plastic products. Traditional feature analysis may not be able to effectively enhance features related to defects, resulting in low accuracy in defect detection. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the disadvantages of the above-mentioned prior art and provide a data processing system based on plastic product detection.

[0005] The technical solution adopted to solve the above technical problems is: a data processing system based on plastic product detection, including an image processing unit, a curved surface detection unit, a defect mapping unit and a plane detection unit, and also includes:

[0006] An image acquisition unit, wherein the image acquisition unit is used to acquire an image of a target plastic product based on an industrial camera at a fixed position to obtain an image of the target plastic product, and transmit the image of the target plastic product to a target detection unit;

[0007] A target detection unit, which receives the target plastic product image transmitted by the image acquisition unit, and performs target detection on the target plastic product image based on a trained target detection model to obtain a curved plastic product sub-image and a flat plastic product sub-image in the target plastic product image, transmits the curved plastic product sub-image to a contour extraction unit, and transmits the flat plastic product sub-image to an image processing unit;

[0008] A contour extraction unit receives the curved plastic product sub-image transmitted by the target detection unit, and constructs a first contour point set according to the trend of a preset contour line, extracts the contour line of the curved plastic product sub-image based on the first contour point set to obtain a contour line image corresponding to the curved plastic product sub-image, and transmits the contour line image to the curved surface detection unit.

[0009] Preferably, the surface detection unit receives the contour line image transmitted by the contour extraction unit, obtains the center point of the contour line image, and calculates the direction vector from the center point to each point in the contour line image to obtain a direction vector set, and performs surface detection on the contour line image based on the direction vector set to obtain a surface contour quality label corresponding to the contour line image, that is, obtains the surface contour quality label corresponding to the curved plastic product sub-image, and transmits the surface contour quality label corresponding to the curved plastic product sub-image to the defect mapping unit.

[0010] Preferably, the image processing unit receives the planar plastic product sub-image transmitted by the target detection unit, and decomposes the planar plastic product sub-image into a first mapping image and a second mapping image based on a trigonometric function decomposition technique, performs high-pass filtering on the first mapping image to obtain a first standard mapping image, performs high-pass filtering on the second mapping image to obtain a second standard mapping image, reconstructs the first standard mapping image and the second standard mapping image based on an inverse trigonometric function decomposition technique to obtain a standard planar plastic product sub-image, and transmits the standard planar plastic product sub-image to the plane detection unit.

[0011] Preferably, the plane detection unit receives the standard plane plastic product sub-image transmitted by the image processing unit, and performs defect detection on the standard plane plastic product sub-image based on a trained defect detection model to obtain a plane defect label corresponding to the standard plane plastic product sub-image, and transmits the plane defect label corresponding to the standard plane plastic product sub-image to the defect mapping unit.

[0012] Preferably, the defect mapping unit receives the surface contour quality label corresponding to the curved plastic product sub-image transmitted by the surface detection unit and the plane defect label corresponding to the standard plane plastic product sub-image transmitted by the plane detection unit, and encodes and assigns values ​​to the surface contour quality label and the plane defect label based on a preset label coding set to obtain a label coding sequence, and matches the label coding sequence with a preset label sequence table to obtain the detection label corresponding to the target plastic product image, that is, to obtain the detection label corresponding to the target plastic product.

[0013] Preferably, extracting the contour line of the curved plastic product sub-image based on the first contour point set to obtain a contour line image corresponding to the curved plastic product sub-image includes:

[0014] The first contour point set The minimum point of the horizontal coordinate of each data is marked and recorded as , the As the first point of the contour;

[0015] For the first contour point set, the rest points to traverse and calculate the remaining Points and The cosine value of the angle in the horizontal direction of the connecting line is obtained to obtain a set of cosine values ​​of the angle;

[0016] The minimum angle cosine value is obtained from the angle cosine value set, and the point corresponding to the minimum angle cosine value is obtained and marked as , the As the second point of the contour line;

[0017] For the first contour point set, the rest points to traverse and construct the rest Points and and stated The triangles formed by the triangles are obtained to obtain a triangle set, and the points corresponding to the triangles with the largest area in the triangle set are obtained and marked as , the As the third point of the contour line;

[0018] For the first contour point set, the rest points to traverse and As new , the As new , and build the rest Points and new and new The triangles formed by the triangles are obtained to obtain a triangle set, and the points corresponding to the triangles with the largest area in the triangle set are obtained and marked as new , repeat the above operation until all points in the first contour point set are traversed to obtain the contour line image corresponding to the curved plastic product sub-image.

[0019] Preferably, performing surface detection on the contour line image based on the direction vector set to obtain a surface contour quality label corresponding to the contour line image includes:

[0020] Adjacent angles in the direction vector set are calculated in sequence to obtain an angle set corresponding to the direction vector set, wherein the angle calculation formula is as follows:

[0021] ;

[0022] in, Indicates the direction vector set and The angle between the direction vectors, Direction vector concentration direction vector;

[0023] Compare the angle set with the preset standard angle set one by one to obtain the absolute value of the difference between each angle in the angle set and each standard angle in the preset standard angle set;

[0024] The absolute values ​​of the differences are summed to obtain a surface detection coefficient, and the surface detection coefficient is compared with a preset surface detection coefficient threshold to obtain a surface contour quality label corresponding to the contour line image.

[0025] Preferably, the expression of the first mapping image is as follows:

[0026] ;

[0027] in, represents the first mapping image, Represents a sub-image of a flat plastic product;

[0028] The expression of the second mapping image is as follows:

[0029] ;

[0030] in, represents a second mapped image;

[0031] The reconstruction expressions of the first standard mapping image and the second standard mapping are as follows:

[0032] ;

[0033] in, Represents a standard flat plastic product sub-image.

[0034] Preferably, the training method of the target detection model includes:

[0035] Based on the labeled samples, the primary target detection model is trained to obtain a trained primary target detection model, where the labeled samples are represented as ;

[0036] The sample without labeled information is input into the trained primary target detection model, and the trained primary target detection model outputs the pseudo label corresponding to the sample without labeled information to obtain the sample with pseudo label, wherein the sample with pseudo label is represented as ;

[0037] The target detection model is trained based on the samples with pseudo labels and the samples with labels to obtain a trained target detection model.

[0038] Preferably, the defect detection model includes a backbone module, a feature fusion module and a multi-scale detection module. The backbone module is used to first input a channel number The input feature map is After the standard convolution is used for dimensionality reduction, the number of channels is Feature map , and then go through The number of channels obtained after depth-wise separable convolution is Feature map , in the feature map Then embed the channel attention mechanism and transform the feature map and feature map The output features are spliced ​​to obtain the output features. The feature fusion module is used to extract feature information of different sizes from the backbone module, complete the first feature fusion through horizontal connection with the downsampling layer, and complete the second feature fusion through jump connection with the same size upsampling layer and downsampling, and finally obtain a feature map of multi-scale feature fusion.

[0039] Preferably, the expression of the backbone module is as follows:

[0040] ;

[0041] in, represents the output features, Indicates disrupting operation. represents the channel attention mechanism, Represents a splicing operation, represents the standard convolution operation, represents the input feature map, Represents a depthwise separable convolution operation.

[0042] The beneficial effects of the present invention are as follows: (1) The present invention detects the collected target plastic product image based on the trained target detection model through the target detection unit, and can accurately separate the curved plastic product sub-image and the flat plastic product sub-image therein, which is helpful for the subsequent use of different detection methods for different types of plastic products, avoiding the inaccuracy that may be caused by unified detection, and making the detection more targeted; (2) For the curved plastic product, the present invention constructs a contour point set and extracts a contour line image according to a preset contour line trend by the contour extraction unit, and then the surface detection unit performs surface detection based on the center point and direction vector set of the contour line image, so as to accurately analyze the surface contour. Quality, for flat plastic products, the image processing unit uses trigonometric function decomposition technology, high-pass filtering and inverse trigonometric function decomposition technology for processing, and then the plane detection unit performs defect detection based on the defect detection model. This targeted detection can deeply analyze the characteristics of different types of plastic products and improve the accuracy and effectiveness of detection; (3) The present invention uses a defect mapping unit to encode and assign surface contour quality labels and plane defect labels based on a preset label coding set, and then matches the obtained label coding sequence with a preset label sequence table to obtain the final detection label. This method standardizes different types of detection results so that the detection results have a unified format and a clear meaning. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic diagram of the system architecture of an overall system in an embodiment of the present invention.

[0044] Figure numerals: 1. Image acquisition unit; 2. Target detection unit; 3. Contour extraction unit; 4. Image processing unit; 5. Surface detection unit; 6. Defect mapping unit; 7. Plane detection unit. DETAILED DESCRIPTION

[0045] Embodiment 1, as Figure 1 As shown, a data processing system based on plastic product detection proposed by the present invention includes an image processing unit 4, a curved surface detection unit 5, a defect mapping unit 6 and a plane detection unit 7, and also includes:

[0046] An image acquisition unit 1 is used to acquire an image of a target plastic product based on an industrial camera at a fixed position to obtain an image of the target plastic product, and transmit the image of the target plastic product to a target detection unit 2;

[0047] The target detection unit 2 receives the target plastic product image transmitted by the image acquisition unit 1, and performs target detection on the target plastic product image based on the trained target detection model to obtain a curved plastic product sub-image and a flat plastic product sub-image in the target plastic product image, transmits the curved plastic product sub-image to the contour extraction unit 3, and transmits the flat plastic product sub-image to the image processing unit 4;

[0048] The contour extraction unit 3 receives the curved plastic product sub-image transmitted by the target detection unit 2, and constructs a first contour point set according to the trend of a preset contour line, extracts the contour line of the curved plastic product sub-image based on the first contour point set to obtain a contour line image corresponding to the curved plastic product sub-image, and transmits the contour line image to the curved surface detection unit 5.

[0049] In the present invention, an industrial camera is an imaging device specially designed for industrial applications. Compared with ordinary cameras, it has higher image quality, stability and reliability. In industrial environments, it is necessary to accurately capture images of various objects. Industrial cameras can usually provide higher resolution, better contrast and lower noise levels; fixed camera position means that the position of the industrial camera when capturing images is fixed, which means that the camera's shooting angle, height, distance and other parameters are all predetermined. The advantage of using a fixed camera position is that the consistency of each captured image can be guaranteed, which is convenient for subsequent image processing and analysis; the target detection model is a model built based on machine learning or deep learning algorithms, which is used to identify and locate specific target objects in images. In this system, it is used to detect curved plastic products and flat plastic products in target plastic product images; the contour point set is a set of points, which are arranged in a certain order in the image, and these points are connected to approximately represent the contour of the target object. In this system, the first contour point set is constructed according to the trend of the preset contour line, that is, according to the prior knowledge of the contour line of the curved plastic product or a certain preset rule, a set of initial points is determined, and these points will serve as the basis for subsequent contour extraction.

[0050] In an optional embodiment, the surface detection unit 5 receives the contour line image transmitted by the contour extraction unit 3, obtains the center point of the contour line image, and calculates the direction vector from the center point to each point in the contour line image to obtain a direction vector set, and performs surface detection on the contour line image based on the direction vector set to obtain a surface contour quality label corresponding to the contour line image, that is, obtains a surface contour quality label corresponding to the curved plastic product sub-image, and transmits the surface contour quality label corresponding to the curved plastic product sub-image to the defect mapping unit 6.

[0051] It should be noted that surface detection is the process of detecting objects with curved shapes (here, curved plastic products, indirectly represented by their contour line images). In this process, the quality of the surface is evaluated by analyzing various features related to the surface. Here, the contour line image is used for surface detection using a direction vector set. The quality of the surface may be judged based on the difference between the shape characteristics of the contour line reflected by the direction vector set and the shape characteristics of the ideal surface.

[0052] In an optional embodiment, the image processing unit 4 receives the planar plastic product sub-image transmitted by the target detection unit 2, and decomposes the planar plastic product sub-image into a first mapping image and a second mapping image based on a trigonometric function decomposition technique, performs high-pass filtering on the first mapping image to obtain a first standard mapping image, performs high-pass filtering on the second mapping image to obtain a second standard mapping image, reconstructs the first standard mapping image and the second standard mapping image based on an inverse trigonometric function decomposition technique to obtain a standard planar plastic product sub-image, and transmits the standard planar plastic product sub-image to the plane detection unit 7.

[0053] It should be noted that trigonometric function decomposition technology is a method of decomposing an image according to the characteristics of trigonometric functions (such as sine function and cosine function). In image processing, an image can be regarded as a two-dimensional signal. Through trigonometric function decomposition technology, this two-dimensional signal can be represented as a combination of trigonometric functions of different frequencies. This decomposition method helps to separate the different frequency components in the image, just like decomposing the signal into a superposition of sine and cosine waves of different frequencies in Fourier transform. In this system, it is used to decompose the plane plastic product sub-image into a first mapping image and a second mapping image. These two mapping images may represent the performance of the original image under different trigonometric function frequency components respectively; high-pass filtering is an operation of filtering a signal (here is the image signal) in the frequency domain. In the frequency domain representation of the image, the image information is decomposed into components of different frequencies, including low-frequency components and high-frequency components.

[0054] In an optional embodiment, the plane detection unit 7 receives the standard plane plastic product sub-image transmitted by the image processing unit 4, and performs defect detection on the standard plane plastic product sub-image based on a trained defect detection model to obtain a plane defect label corresponding to the standard plane plastic product sub-image, and transmits the plane defect label corresponding to the standard plane plastic product sub-image to the defect mapping unit 6.

[0055] In an optional embodiment, the defect mapping unit 6 receives the surface contour quality label corresponding to the curved plastic product sub-image transmitted by the surface detection unit 5 and the plane defect label corresponding to the standard plane plastic product sub-image transmitted by the plane detection unit 7, and encodes and assigns the surface contour quality label and the plane defect label based on a preset label coding set to obtain a label coding sequence, and matches the label coding sequence with a preset label sequence table to obtain a detection label corresponding to the target plastic product image, that is, to obtain a detection label corresponding to the target plastic product.

[0056] It should be noted that the label coding set is a predefined set, which contains the rules and corresponding values ​​for encoding different labels (here, surface profile quality labels and plane defect labels). This set provides a basis for converting various quality labels into digital codes that can be processed by computers. For example, for surface profile quality labels, if there are three cases of "excellent", "good" and "poor", the label coding set may stipulate that "excellent" corresponds to code 1, "good" corresponds to code 2, and "poor" corresponds to code 3; there is a similar coding correspondence for plane defect labels; coding assignment refers to the operation of assigning corresponding digital codes to surface profile quality labels and plane defect labels according to the rules of the label coding set. Through this operation, labels originally expressed in text or other forms are converted into digital form. For example, the surface profile quality label of a curved plastic product is "good", and according to the provisions of the label coding set, it will be assigned a value of 2; if the plane defect label is "slightly Defects" will also be assigned corresponding digital codes, such as 4 (assuming that there is such a provision in the label coding set); the label coding sequence is a sequence formed by arranging the surface profile quality label codes and plane defect label codes after coding assignment in a certain order. For example, if the surface profile quality label is first coded and assigned to code 2, and then the plane defect label is coded and assigned to code 4, then the label coding sequence may be [2,4]. This sequence integrates the coding information of the target plastic product in terms of surface profile quality and plane defects; the inspection label is an identification of the final inspection result of the target plastic product. It is obtained by matching the label coding sequence with the label sequence table. This label can comprehensively reflect the inspection status of the target plastic product in terms of surface profile quality, plane defects, etc., such as different quality grade identifications such as "high-quality products", "qualified products" and "unqualified products", so as to facilitate the classification, quality control and decision-making of plastic products.

[0057] Embodiment 2, a data processing system based on plastic product detection proposed by the present invention, compared with embodiment 1, this embodiment further includes: extracting the contour line of the curved plastic product sub-image based on the first contour point set to obtain a contour line image corresponding to the curved plastic product sub-image, including:

[0058] The first contour point is concentrated The minimum point of the horizontal coordinate of each data is marked and recorded as ,Will As the first point of the contour;

[0059] For the rest of the first contour points points to traverse and calculate the remaining Points and The cosine value of the angle in the horizontal direction of the connecting line is obtained to obtain a set of cosine values ​​of the angle;

[0060] Get the minimum angle cosine value from the angle cosine value set, get the point corresponding to the minimum angle cosine value, and mark it as ,Will As the second point of the contour line;

[0061] For the rest of the first contour points points to traverse and construct the rest Points and and The triangles formed by the triangles are obtained to obtain a triangle set, and the points corresponding to the triangles with the largest area in the triangle set are obtained and marked as ,Will As the third point of the contour line;

[0062] For the rest of the first contour points points to traverse, As new ,Will As new , and build the rest Points and new and new The triangles formed by the triangles are obtained to obtain a triangle set, and the points corresponding to the triangles with the largest area in the triangle set are obtained and marked as new , repeat the above operation until all points in the first contour point set are traversed to obtain the contour line image corresponding to the curved plastic product sub-image.

[0063] In an optional embodiment, performing surface detection on the contour line image based on the direction vector set to obtain a surface contour quality label corresponding to the contour line image includes:

[0064] The adjacent angles in the direction vector set are calculated in sequence to obtain the angle set corresponding to the direction vector set, where the angle calculation formula is as follows:

[0065] ;

[0066] in, Indicates the direction vector set and The angle between the direction vectors, Direction vector concentration direction vector;

[0067] The angle set is compared with the preset standard angle set one by one to obtain the absolute value of the difference between each angle in the angle set and each standard angle in the preset standard angle set;

[0068] The absolute values ​​of the differences are summed to obtain a surface detection coefficient, and the surface detection coefficient is compared with a preset surface detection coefficient threshold to obtain a surface contour quality label corresponding to the contour line image.

[0069] In an optional embodiment, the expression of the first mapping image is as follows:

[0070] ;

[0071] in, represents the first mapping image, Represents a sub-image of a flat plastic product;

[0072] The expression of the second mapping image is as follows:

[0073] ;

[0074] in, represents a second mapped image;

[0075] The reconstruction expressions of the first standard mapping image and the second standard mapping are as follows:

[0076] ;

[0077] in, Represents a standard flat plastic product sub-image.

[0078] In an optional embodiment, the training method of the target detection model includes:

[0079] Based on the labeled samples, the primary target detection model is trained to obtain a trained primary target detection model, where the labeled samples are represented as ;

[0080] The samples without labeled information are input into the trained primary target detection model, and the trained primary target detection model outputs the pseudo labels corresponding to the samples without labeled information to obtain the samples with pseudo labels, where the samples with pseudo labels are represented as ;

[0081] The target detection model is trained based on the samples with pseudo labels and the samples with labels to obtain a trained target detection model.

[0082] In an optional embodiment, the defect detection model includes a backbone module, a feature fusion module and a multi-scale detection module. The backbone module is used to first input a channel number The input feature map is After the standard convolution is used for dimensionality reduction, the number of channels is Feature map , and then go through The number of channels obtained after depth-wise separable convolution is Feature map , in the feature map Then embed the channel attention mechanism and transform the feature map and feature map The splicing is performed to obtain the output features. The feature fusion module is used to extract feature information of different sizes from the backbone module, complete the first feature fusion through horizontal connection with the downsampling layer, and complete the second feature fusion through jump connection with the upsampling layer and downsampling of the same size, and finally obtain the feature map of multi-scale feature fusion.

[0083] In an optional embodiment, the expression of the backbone module is as follows:

[0084] ;

[0085] in, represents the output features, Indicates disrupting operation. represents the channel attention mechanism, Represents a splicing operation, represents the standard convolution operation, represents the input feature map, Represents a depthwise separable convolution operation.

[0086] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A data processing system based on plastic product detection, comprising an image processing unit (4), a curved surface detection unit (5), a defect mapping unit (6) and a plane detection unit (7), characterized in that: An image acquisition unit (1), the image acquisition unit (1) being used to acquire an image of a target plastic product based on an industrial camera at a fixed position to obtain an image of the target plastic product, and to transmit the image of the target plastic product to a target detection unit (2); A target detection unit (2), wherein the target detection unit (2) receives the target plastic product image transmitted by the image acquisition unit (1), and performs target detection on the target plastic product image based on a trained target detection model to obtain a curved plastic product sub-image and a flat plastic product sub-image in the target plastic product image, transmits the curved plastic product sub-image to the contour extraction unit (3), and transmits the flat plastic product sub-image to the image processing unit (4); A contour extraction unit (3), the contour extraction unit (3) receiving the curved plastic product sub-image transmitted by the target detection unit (2), and constructing a first contour point set according to a trend of a preset contour line, extracting the contour line of the curved plastic product sub-image based on the first contour point set to obtain a contour line image corresponding to the curved plastic product sub-image, and transmitting the contour line image to the curved surface detection unit (5); The surface detection unit (5) receives the contour line image transmitted by the contour extraction unit (3), obtains the center point of the contour line image, and calculates the direction vector from the center point to each point in the contour line image to obtain a direction vector set, performs surface detection on the contour line image based on the direction vector set to obtain a surface contour quality label corresponding to the contour line image, that is, obtains a surface contour quality label corresponding to the curved surface plastic product sub-image, and transmits the surface contour quality label corresponding to the curved surface plastic product sub-image to the defect mapping unit (6); The image processing unit (4) receives the plane plastic product sub-image transmitted by the target detection unit (2), decomposes the plane plastic product sub-image into a first mapping image and a second mapping image based on a trigonometric function decomposition technique, performs high-pass filtering on the first mapping image to obtain a first standard mapping image, performs high-pass filtering on the second mapping image to obtain a second standard mapping image, reconstructs the first standard mapping image and the second standard mapping image based on an inverse trigonometric function decomposition technique to obtain a standard plane plastic product sub-image, and transmits the standard plane plastic product sub-image to the plane detection unit (7); Extracting the contour line of the curved plastic product sub-image based on the first contour point set to obtain a contour line image corresponding to the curved plastic product sub-image includes: The first contour point set The minimum point of the horizontal coordinate of each data is marked and recorded as , the As the first point of the contour; For the first contour point set, the rest points to traverse and calculate the remaining Points and The cosine value of the angle in the horizontal direction of the connecting line is obtained to obtain a set of cosine values ​​of the angle; The minimum angle cosine value is obtained from the angle cosine value set, and the point corresponding to the minimum angle cosine value is obtained and marked as , the As the second point of the contour line; For the first contour point set, the rest points to traverse and construct the rest Points and and stated The triangles formed by the triangles are obtained to obtain a triangle set, and the points corresponding to the triangles with the largest area in the triangle set are obtained and marked as , the As the third point of the contour line; For the first contour point set, the rest points to traverse and As new , the As new , and build the rest Points and new and new The triangles formed by the triangles are obtained to obtain a triangle set, and the points corresponding to the triangles with the largest area in the triangle set are obtained and marked as new , repeat the above operation until all points in the first contour point set are traversed to obtain the contour line image corresponding to the curved plastic product sub-image; Performing surface detection on the contour line image based on the direction vector set to obtain a surface contour quality label corresponding to the contour line image includes: Adjacent angles in the direction vector set are calculated in sequence to obtain an angle set corresponding to the direction vector set, wherein the angle calculation formula is as follows: ； in, Indicates the direction vector set and The angle between the direction vectors, Indicates the direction vector set direction vector; Compare the angle set with the preset standard angle set one by one to obtain the absolute value of the difference between each angle in the angle set and each standard angle in the preset standard angle set; summing the absolute values ​​of the differences to obtain a surface detection coefficient, and comparing the surface detection coefficient with a preset surface detection coefficient threshold to obtain a surface contour quality label corresponding to the contour line image; The training method of the target detection model includes: Based on the labeled samples, the primary target detection model is trained to obtain a trained primary target detection model, where the labeled samples are represented as ; The sample without labeled information is input into the trained primary target detection model, and the trained primary target detection model outputs the pseudo label corresponding to the sample without labeled information to obtain the sample with pseudo label, wherein the sample with pseudo label is represented as ; The target detection model is trained based on the samples with pseudo labels and the samples with labels to obtain a trained target detection model.

2. A data processing system based on plastic product detection according to claim 1, characterized in that: The plane detection unit (7) receives the standard plane plastic product sub-image transmitted by the image processing unit (4), and performs defect detection on the standard plane plastic product sub-image based on the trained defect detection model to obtain a plane defect label corresponding to the standard plane plastic product sub-image, and transmits the plane defect label corresponding to the standard plane plastic product sub-image to the defect mapping unit (6); The defect mapping unit (6) receives the surface contour quality label corresponding to the curved plastic product sub-image transmitted by the surface detection unit (5) and the plane defect label corresponding to the standard plane plastic product sub-image transmitted by the plane detection unit (7), and encodes and assigns values ​​to the surface contour quality label and the plane defect label based on a preset label encoding set to obtain a label encoding sequence, and matches the label encoding sequence with a preset label sequence table to obtain a detection label corresponding to the target plastic product image, that is, to obtain a detection label corresponding to the target plastic product.

3. A data processing system based on plastic product detection according to claim 2, characterized in that: The expression of the first mapping image is as follows: ； in, represents the first mapping image, Represents a sub-image of a flat plastic product; The expression of the second mapping image is as follows: ； in, represents a second mapped image; The reconstruction expressions of the first standard mapping image and the second standard mapping are as follows: ； in, Represents a standard flat plastic product sub-image.

4. A data processing system based on plastic product detection according to claim 3, characterized in that: The defect detection model includes a backbone module, a feature fusion module and a multi-scale detection module. The backbone module is used to first input a channel number of The input feature map is After the standard convolution is used for dimensionality reduction, the number of channels is Feature map , and then go through The number of channels obtained after depth-wise separable convolution is Feature map , in the feature map Then embed the channel attention mechanism and transform the feature map and feature map The output features are spliced ​​to obtain the output features. The feature fusion module is used to extract feature information of different sizes from the backbone module, complete the first feature fusion through horizontal connection with the downsampling layer, and complete the second feature fusion through jump connection with the same size upsampling layer and downsampling, and finally obtain a feature map of multi-scale feature fusion.

5. A data processing system based on plastic product detection according to claim 4, characterized in that: The expression of the backbone module is as follows: ； in, represents the output features, Indicates disrupting operation. represents the channel attention mechanism, Represents a splicing operation, represents the standard convolution operation, represents the input feature map, Represents a depthwise separable convolution operation.

Images