An intelligent defect identification and correction method and system based on interactive target detection

Abstract

This invention belongs to the fields of intelligent manufacturing and visual recognition technology, and specifically relates to an intelligent defect identification and correction method and system based on interactive target detection. The method processes images of specific steps in the manufacturing process using an industrial camera, identifies manufacturing defects in those steps, and communicates with a lower-level machine for error correction. First, the image of the specific step is input into a target detection model based on a convolutional neural network. Even under complex conditions, such as when the target has high transparency or a small shape, the model can still accurately identify whether the target has been correctly processed. Then, the identified results are stored in a structured video format and fed back to the lower-level machine execution mechanism to achieve subsequent processing such as automatic reprocessing, scrapping, and manual intervention. Furthermore, to improve the model's generalization and versatility, the method allows for interactive information input to modify the inference method of the target detection model, thereby affecting the inference results.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and visual recognition technology, specifically to an intelligent defect identification and correction method and system based on interactive target detection. Background Technology

[0002] In the fields of intelligent manufacturing and industrial automation, real-time quality monitoring and defect identification during the production process are crucial for ensuring product pass rates and improving production efficiency. Traditional defect detection methods mainly rely on manual visual inspection or machine vision systems based on fixed rules. Manual inspection suffers from low efficiency, fatigue, strong subjectivity, and poor consistency, making it difficult to meet the demands of modern high-speed, high-precision production. While machine vision systems based on traditional image processing (such as threshold segmentation, edge detection, and template matching) have achieved automation to some extent, their performance is heavily dependent on pre-set rules and idealized imaging conditions. When faced with complex and ever-changing real-world industrial scenarios, such as targets with high transparency, small structures, varied colors, or severe background interference, the robustness and adaptability of traditional methods decrease significantly, easily leading to false positives or false negatives.

[0003] In recent years, breakthroughs in deep learning technology, especially object detection algorithms based on convolutional neural networks (CNNs) (such as Faster R-CNN and the YOLO series), have demonstrated powerful performance in computer vision tasks. These models can automatically learn deep features in images, and their detection accuracy for common objects on various public datasets has surpassed that of traditional methods. Therefore, some advanced manufacturing systems have begun to explore the introduction of deep learning models for defect detection.

[0004] However, directly applying existing deep learning object detection models to industrial defect identification and correction scenarios still faces several technical bottlenecks: First, the morphology, size, and characteristics of industrial defects vary greatly depending on the product, process, and material physical parameters (such as color, transparency, thickness, and reflectivity). A fixed model trained on a specific dataset has limited generalization ability. When the production line changes product models or material properties, the model's performance may deteriorate sharply, requiring data collection and model retraining, resulting in long system downtime and high update and maintenance costs. Second, most existing detection systems are "open-loop" designs, meaning the detection model is an independent judgment module with fixed inference logic, unable to be dynamically adjusted based on engineers' domain knowledge or real-time feedback. When encountering situations with low model confidence or near the judgment boundary, the system lacks a flexible human-computer interaction mechanism to correct the inference process or results, potentially leading to incorrect decisions. Finally, most studies focus on simple defect identification and fail to deeply integrate the identification results with downstream production execution mechanisms (lower-level machines) to form a closed-loop control system of "perception-decision-execution". As a result, true automated error correction, such as automatic guidance for reprocessing, classification and rejection, or precise prompts for human intervention, cannot be achieved.

[0005] Therefore, there is an urgent need in the existing technology to develop an intelligent defect identification and correction method that has strong adaptability to complex working conditions, can flexibly integrate human expert experience, and can control the actuator in a closed loop, so as to improve the accuracy, robustness, versatility and automation level of intelligent manufacturing quality inspection system. To this end, we propose an intelligent defect identification and correction method and system based on interactive target detection to solve the above problems. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent defect identification and correction method and system based on interactive target detection. This application introduces a target detection model based on convolutional neural networks and combines it with an interactive information input mechanism to accurately identify the state of the object being tested, thereby achieving real-time detection, structured storage, and feedback control of defects. This improves the automation level, detection stability, and quality controllability of intelligent manufacturing production lines, and solves the problems mentioned in the background technology.

[0007] (II) Technical Solution To achieve the above objectives, the present invention specifically adopts the following technical solution: An intelligent defect identification and correction method based on interactive target detection includes the following steps: S1. Image Acquisition Steps: Acquire image data or video stream data of specific processes in the manufacturing process using an industrial camera; S2. Interactive reasoning and recognition step: The collected specific process images are combined with interactive information and input into the target detection model for reasoning and recognition. The target detection model is a target detection model based on a convolutional neural network, which is used to identify whether the object under test has correctly achieved the process target. S3. Defect Judgment Step: Based on the reasoning and recognition results of the target detection model described in step S2, determine whether the object under test has manufacturing defects; S4. Structured storage step: The reasoning and identification results of step S2 and the defect judgment results of step S3 are converted into structured video data for storage, and a structured record containing the state information of the object under test is generated to achieve traceable management of the manufacturing process. S5. Closed-loop control step: The status information of the object under test generated in step S4 is sent to the actuator or lower control system linked to the production line to trigger the processing action corresponding to the status information. The processing action includes automatic reprocessing, scrap disposal or requesting manual intervention to realize closed-loop control of product quality in the manufacturing process.

[0008] Furthermore, the target detection model is a multi-stage target detection model based on convolutional neural networks to improve the detection accuracy of small objects and facilitate the incorporation of interactive information. The inference process of the multi-stage target detection model includes: an image preprocessing layer, an image feature extraction layer, a linear layer, and a post-processing layer for inference results. The image preprocessing layer typically includes operations such as image cropping, image scaling, and image transformation to improve image feature discrimination and neural network recognition. The image feature extraction layer uses a convolutional neural network to extract image features related to defects, facilitating subsequent probability induction. The linear layer inductively summarizes the image features to output the probability and location of each defect. The post-processing layer further optimizes the results of the linear layer to eliminate possible false positives. The inference results of the target detection model include one or more of the following: process completed, process not completed, no object detected, and abnormal object location.

[0009] Furthermore, the target detection model can accurately detect objects in complex working scenarios, such as those with high transparency and small size.

[0010] Furthermore, the interactive information is used to modify the inference method of the object detection model. This modification includes pre-inference modification, in-inference modification, and post-inference modification. Pre-inference modification refers to the image preprocessing layer in the multi-stage object detection model modifying the image transformation based on the object color information and transparency parameters in the interactive information. This ensures that the actual input image during deployment is as close as possible to the distribution of the multi-stage object detection model's training set, thereby improving the model's generalization and robustness. In-inference modification refers to performing gradient backpropagation on the linear layer in the multi-stage object detection model based on the defect-free image, the defective image, and detection box information, thereby fine-tuning the linear layer parameters to achieve accurate identification of newly emerging defects. Post-inference modification refers to adjusting the parameters of the inference result post-processing layer in the multi-stage object detection model based on the defective image and detection box information.

[0011] Furthermore, the interactive information is input through a human-computer interaction interface or a host system, and affects the inference results of the target detection model without retraining the complete target detection model, thereby improving the model's adaptability to different production scenarios.

[0012] Furthermore, in step S4, the video structured data includes at least one or more of the following: process sequence number, process time information, location, detection result, defect type, and image or video index information.

[0013] Furthermore, in step S5, the actuator includes an execution device, a correction device, or an automated control unit that is communicatively connected thereto.

[0014] The present invention also provides an intelligent defect identification and error correction system based on interactive target detection, including an industrial camera module, a processor and a memory.

[0015] The industrial camera module is used to acquire image or video data of specific processes in the manufacturing process; the memory stores a computer program; and the processor, when executing the computer program, is used to implement the defect identification method steps described in any of the above method schemes.

[0016] (III) Beneficial Effects Compared with existing technologies, this invention provides an intelligent defect identification and correction method and system based on interactive target detection, which has the following beneficial effects: 1. Improve defect detection accuracy: By using a target detection model based on convolutional neural networks, the problem of complex backgrounds, high transparency, and small targets of the tested objects can be effectively overcome, thereby improving the accuracy and stability of defect identification.

[0017] 2. Introduce an interactive reasoning mechanism to enhance the model's generalization ability: The model's reasoning method is dynamically adjusted through interactive information input, adapting to different physical properties of the tested object, such as changes in color, transparency, and thickness, without retraining the complete model.

[0018] 3. Achieve closed-loop control of detection and feedback: Feed back the detection results to the actuator or lower control system in real time to realize automatic remanufacturing, scrapping or manual intervention, thereby improving the automation level and quality control capability of the production line.

[0019] 4. Supports structured video storage and quality traceability: The test results are stored in a structured video format, which facilitates subsequent quality analysis, problem tracing, and production optimization.

[0020] 5. The system has a clear structure and is easy to deploy in engineering: the methods and systems correspond to each other, making it easy to integrate and deploy in existing smart manufacturing production lines, and it has high industrial application value. Attached Figure Description

[0021] Figure 1 This is the overall process of an intelligent defect identification and correction method based on interactive target detection. Detailed Implementation

[0022] 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.

[0023] like Figure 1 As shown in the figure, an intelligent defect identification and correction method based on interactive target detection proposed in one embodiment of the present invention includes the following steps: S1. Image Acquisition Steps: Acquire image data or video stream data of specific processes in the manufacturing process using an industrial camera; S2, Interactive Reasoning and Recognition Step: The image data or video stream data collected in step S1 is combined with the externally input interactive information and input into the target detection model based on convolutional neural network for reasoning and recognition to determine whether the target object has correctly completed the specific process; the interactive information is used to dynamically adjust the reasoning process of the target detection model without retraining the complete target detection model.

[0024] The target detection model is a multi-stage target detection model based on a convolutional neural network. The inference process of the multi-stage target detection model includes: an image preprocessing layer, an image feature extraction layer, a linear layer, and an inference result post-processing layer, specifically including: Image preprocessing stage: Perform one or more preprocessing operations on the input image data, including cropping, scaling, normalization, or color space transformation, to obtain the preprocessed image; Image feature extraction stage: Deep feature extraction is performed on the preprocessed image using a convolutional neural network to obtain a high-dimensional feature map representing the state of the object under test; Feature induction and prediction stage: The high-dimensional feature map is inductively mapped through at least one linear layer to output the state category probability of the object under test and its position information in the image; Post-processing stage of inference results: The results output from the feature summarization and prediction stage are subjected to confidence screening and / or non-maximum suppression processing to obtain the final set of detection results.

[0025] The target detection model can accurately detect objects in complex working scenarios, such as those with high transparency and small size.

[0026] S3. Defect Judgment Step: Based on the reasoning and recognition results of the target detection model described in step S2, determine whether the object under test has manufacturing defects; S4. Structured storage step: Convert the reasoning and recognition results of step S2 and the defect judgment results of step S3 into structured video data for storage, and generate a structured record containing the state information of the object under test. S5. Closed-loop control step: Send the status information of the object under test generated in step S4 to the actuator or lower control system linked to the production line to trigger the processing action corresponding to the status information. The processing action includes automatic reprocessing, scrap disposal, or requesting manual intervention.

[0027] This method processes images of specific steps in the manufacturing process using industrial cameras to identify manufacturing defects and communicate with a lower-level machine for error correction. First, the image of the specific step is input into a target detection model based on a convolutional neural network. Even under complex conditions, such as when the target has high transparency or a small shape, the model can still accurately identify whether the target has been correctly processed. Then, the identified results are stored in a structured video format and fed back to the lower-level execution mechanism for subsequent processing such as automatic reprocessing, scrapping, and manual intervention. Furthermore, to improve the model's generalization and versatility, this method allows for interactive information input to modify the inference method of the target detection model, thereby affecting the inference results to adapt to different physical parameters of the target, such as color, transparency, and thickness.

[0028] like Figure 1As shown, in some embodiments, the interactive information includes the color information of the object under test, the transparency parameter of the object under test, a defect-free image, a defective image, and one or more of the following in the detection frame:

[0029] like Figure 1 As shown, in some embodiments, in step S2, the interactive information is used to modify the inference method of the target detection model. The modification method includes pre-inference modification, in-inference modification, and post-inference modification, specifically including: Pre-inference modification: Based on the physical parameters of the object under test contained in the interaction information, the transformation strategy and parameters in the image preprocessing stage are dynamically adjusted to match the distribution of the input image with the distribution of the model training data; the physical parameters of the object under test include one or more of color information, transparency parameters, and thickness information; Modification during inference: Based on the labeled sample images and corresponding detection box information contained in the interaction information, a supervised loss function is constructed, and the linear layer parameters in the feature induction and prediction stage are fine-tuned online using the gradient backpropagation algorithm; the labeled sample images include defect-free sample images and / or defective sample images; Post-inference modification: Based on the defect sample information contained in the interaction information, dynamically adjust the confidence threshold or result filtering rules in the post-processing stage of the inference results.

[0030] like Figure 1 As shown, in some embodiments, the video structured data in step S4 includes at least one or more of the following: process number, process time information, location, detection result, defect type, and image or video index information.

[0031] like Figure 1 As shown, in some embodiments, the actuator in step S5 includes a manufacturing apparatus, a correction apparatus, or an automated control unit communicatively connected thereto.

[0032] like Figure 1 As shown, in some embodiments, the interactive information is input through a human-computer interaction interface or a host system, and affects the inference results of the target detection model without retraining the complete model.

[0033] like Figure 1 As shown, in some embodiments, the inference results of the target detection model include one or more of the following: Process completed, process not completed, no object detected, object in abnormal position.

[0034] The present invention also provides an intelligent defect identification and error correction system based on interactive target detection, including an industrial camera module, a processor and a memory.

[0035] The industrial camera module is used to acquire image or video data of specific processes in the manufacturing process; the memory stores a computer program; and the processor, when executing the computer program, is used to implement the defect identification method steps described in any of the above method schemes.

[0036] Specific Implementation: The intelligent defect identification and correction method based on interactive target detection described in this invention is applied to an intelligent manufacturing production line for real-time quality inspection and feedback control of products in specific processes. The specific implementation process is as follows: S1. Image Acquisition Step: In step S1, an industrial camera is installed at a specific workstation. The industrial camera is fixedly installed at a predetermined position that can cover the area of ​​the process operation and operates synchronously with the intelligent manufacturing production line. The industrial camera continuously captures images of the production process according to preset parameters such as frame rate and gain, acquiring image or video data of the object under test, and transmitting the acquired data to the back-end processing unit in real time to achieve real-time monitoring of the manufacturing process.

[0037] In this embodiment, the industrial camera can be an area scan industrial camera or a line scan industrial camera, and the acquisition method can be continuous video streaming or triggered image acquisition, with no specific form limited.

[0038] S2, Target Detection and Interactive Reasoning Step: In step S2, the collected process images, combined with interactive information, are input into the target detection model for reasoning and recognition. The target detection model is a multi-stage target detection model based on a convolutional neural network, used to locate and identify the object under test in order to determine whether the object under test is correctly manufactured.

[0039] In this embodiment, the inference process of the multi-stage target detection model includes the following stages: 1. Image Preprocessing Layer: The image preprocessing layer performs preprocessing operations on the input images of specific processes. The preprocessing operations include, but are not limited to, image cropping, image scaling, image normalization, and image transformation, which are used to enhance the feature discrimination between the tested object and the background and improve the recognition of image features by the neural network.

[0040] Let the specific process image captured by the industrial camera be represented as: ; In the formula, H and W are the image height and width, respectively, and C is the number of color channels.

[0041] The image preprocessing layer performs the following transformations on the input image of the specific process: In the formula, T represents the image transformation operator, including cropping, scaling, channel color transformation, etc. The parameters of the image transformation operator include the cropping position (x, y), cropping height and width (H, W), and scaling factor (...). ), Channel color transformation function parameters .

[0042] 2. Image Feature Extraction Layer: The preprocessed image is input to the image feature extraction layer, which extracts multi-level image features related to the defects of the tested object through a convolutional neural network, providing a feature basis for subsequent state judgment of the tested object.

[0043] Image after preprocessing The input is fed into the image feature extraction layer. Let the convolutional neural network's... The layer output feature map is: ; In the formula, ; and The first Layer convolution kernel parameters and biases; " indicates convolution operation; This represents a non-linear activation function.

[0044] Finally, a high-dimensional feature representation related to the defects of the tested object is obtained: in This represents the entire feature extraction network.

[0045] 3. Linear layer: The linear layer summarizes and maps the extracted image features, and outputs the existence probability, spatial location and corresponding state probability of the target object.

[0046] The linear layer represents the features. The results are summarized and output as follows: Predicted state of the measured object and predicted position of the target: ; In the formula, , For linear layer parameters; Output vectors for the model.

[0047] Map the output to probabilities using the Softmax or Sigmoid function: In the formula, This indicates the number of state categories of the object being measured.

[0048] The final set of states of the tested object output by the model is: .

[0049] 4. Post-processing layer for inference results: The post-processing layer for inference results further processes the results output by the linear layer to remove low-confidence detection results or possible false positive results, thereby obtaining the final target detection output.

[0050] The set of detection results output by the linear layer is denoted as: ; In the formula, Indicates the first One detection box; This indicates the corresponding detection confidence level.

[0051] The post-processing layer of inference results is based on a threshold. Perform filtering: The inference results of the target detection model include one or more of the following: process completed, process not completed, no object detected, and object location abnormal.

[0052] S3. Interactive Information Mechanism: In this embodiment, the interactive information includes one or more of the following: the color information of the object under test, the transparency parameter of the object under test, a defect-free image, an image with defects, and a detection frame. The interactive information is input through a human-computer interaction interface or a host system and is used to modify the inference method of the target detection model.

[0053] Specifically, the interactive information can influence the model inference process at the following stages.

[0054] 1. Pre-inference modification: In the pre-inference stage, the image preprocessing layer dynamically adjusts the image transformation strategy based on the color information and transparency parameters of the tested object in the interaction information, so that the distribution of the input image in the actual deployment environment is as close as possible to the image distribution of the model training set, thereby improving the generalization ability and robustness of the model in different production scenarios.

[0055] When the interactive information is introduced, the image transformation operator parameters Based on the color information of the object being measured Transparency parameter of the object under test and other color information of the object being tested Dynamic adjustment: In the formula, To adjust the method, the main adjustment is to the channel color transformation function parameter in the image transformation operator parameters. It can be a piecewise function, an exponential function, or a logarithmic function, etc.

[0056] This allows the actual input image distribution to be as close as possible to the model training data distribution, improving the model's generalization ability under different transparency and color conditions.

[0057] 2. Modification during inference: During the inference process, based on the input defect-free image, defective image and detection box information, gradient backpropagation is performed on the linear layer in the multi-stage target detection model to fine-tune the parameters of the linear layer, thereby improving the model's ability to identify newly emerging defect types.

[0058] Incorporating the defect-free image, the defective image, and the detection box information, a supervised loss function is constructed as follows: In the formula, This represents the loss for classifying the state of the object being measured. This represents the regression loss for the location of the measured object; These are the weighting coefficients.

[0059] Based on gradient backpropagation, only the parameters of the linear layer are updated: In the formula, This is the learning rate.

[0060] This process does not involve updating the parameters of the image feature extraction layer, thus enabling rapid adaptation to newly emerging defects without retraining the complete object detection model.

[0061] 3. Post-inference modification: After inference is completed, the parameters of the post-processing layer of the inference result are adjusted according to the defective image and detection box information to optimize the detection threshold or screening rules and reduce false detections and missed detections.

[0062] After introducing interactive information, the threshold is dynamically adjusted based on the defective image and the detection box information: The interactive modification process described above is completed without retraining the complete object detection model, thus enabling rapid adaptation to changes in parameters of different tested objects.

[0063] S4. Defect Judgment and Data Storage Steps for the Tested Object In step S3, the state of the object under test is analyzed and judged based on the inference results output by the target detection model to determine whether there are defects such as incompleteness, missing parts or abnormal positions.

[0064] In step S4, the reasoning and recognition results are stored in the form of structured video data. The structured video data includes at least one or more of the following: process sequence number, process time information, location, detection result, defect type, and image or video index information, thereby enabling traceable management of the manufacturing process.

[0065] S5, Feedback Control Steps In step S5, the status information of the object under test is fed back to the actuator or lower-level control system. The actuator includes a manufacturing device, a correction device, or an automated control unit that is communicatively connected to it.

[0066] In this embodiment, when the detection result indicates that the tested object has defects, the upper control system can trigger automatic reprocessing, scrapping, and manual intervention based on the status information of the tested object, thereby realizing closed-loop control of product quality.

[0067] In summary, this invention provides an intelligent defect identification and correction method based on interactive target detection, which has the following beneficial effects: Improve defect detection accuracy: By using a target detection model based on convolutional neural networks, the problems of complex backgrounds, high transparency, and small targets can be effectively overcome, thereby improving the accuracy and stability of defect identification.

[0068] An interactive reasoning mechanism is introduced to enhance the model's generalization ability: the model's reasoning method is dynamically adjusted through interactive information input, adapting to different physical properties of the tested object, such as changes in color, transparency, and thickness, without retraining the complete model.

[0069] Achieve closed-loop control of detection and feedback: Feed back the detection results to the actuator or lower control system in real time to realize automatic remanufacturing, scrapping or manual intervention, thereby improving the automation level and quality control capability of the production line.

[0070] Supports structured video storage and quality traceability: The test results are stored in a structured video format, which facilitates subsequent quality analysis, problem tracing, and production optimization.

[0071] The system has a clear structure and is easy to deploy in engineering: the methods and systems correspond to each other, making it easy to integrate and deploy in existing smart manufacturing production lines, and it has high industrial application value.

[0072] 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 foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent defect identification and correction method based on interactive target detection, characterized in that, Includes the following steps: S1. Image Acquisition Steps: Acquire image data or video stream data of specific processes in the manufacturing process using an industrial camera; S2. Interactive reasoning and recognition step: The collected specific process images are combined with interactive information and input into the target detection model for reasoning and recognition. The target detection model is a target detection model based on a convolutional neural network, which is used to identify whether the object under test has correctly achieved the process target. S3. Defect Judgment Step: Based on the reasoning and recognition results of the target detection model described in step S2, determine whether the object under test has manufacturing defects; S4. Structured storage step: Convert the reasoning and recognition results of step S2 and the defect judgment results of step S3 into structured video data for storage, and generate a structured record containing the state information of the object under test. S5. Closed-loop control step: Send the status information of the object under test generated in step S4 to the actuator or lower control system linked to the production line to trigger the processing action corresponding to the status information. The processing action includes automatic reprocessing, scrap disposal, or requesting manual intervention.

2. The intelligent defect identification and correction method based on interactive target detection according to claim 1, characterized in that: The target detection model is a multi-stage target detection model based on convolutional neural networks.

3. The intelligent defect identification and correction method based on interactive target detection according to claim 2, characterized in that: In step S2, the multi-stage target detection model inference process includes: an image preprocessing layer, an image feature extraction layer, a linear layer, and an inference result post-processing layer, specifically including: Image preprocessing stage: Perform one or more preprocessing operations on the input image data, including cropping, scaling, normalization, or color space transformation, to obtain the preprocessed image; Image feature extraction stage: Deep feature extraction is performed on the preprocessed image using a convolutional neural network to obtain a high-dimensional feature map representing the state of the object under test; Feature induction and prediction stage: The high-dimensional feature map is inductively mapped through at least one linear layer to output the state category probability of the object under test and its position information in the image; Post-processing stage of inference results: The results output from the feature summarization and prediction stage are subjected to confidence screening and / or non-maximum suppression processing to obtain the final set of detection results. The target detection model can accurately detect objects in complex working scenarios, such as those with high transparency and small size.

4. The intelligent defect identification and correction method based on interactive target detection according to claim 3, characterized in that: The interactive information includes the color information of the object under test, the transparency parameter of the object under test, a defect-free image, a defective image, and one or more of the following in the detection frame:

5. The intelligent defect identification and correction method based on interactive target detection according to claim 3, characterized in that: In step S2, the interactive information is used to modify the inference method of the target detection model. The modification method includes pre-inference modification, in-inference modification, and post-inference modification, specifically including: Pre-inference modification: Based on the physical parameters of the object under test contained in the interaction information, the transformation strategy and parameters in the image preprocessing stage are dynamically adjusted to match the distribution of the input image with the distribution of the model training data; the physical parameters of the object under test include one or more of color information, transparency parameters, and thickness information; Modification during inference: Based on the labeled sample images and corresponding detection box information contained in the interaction information, a supervised loss function is constructed, and the linear layer parameters in the feature induction and prediction stage are fine-tuned online using the gradient backpropagation algorithm; the labeled sample images include defect-free sample images and / or defective sample images; Post-inference modification: Based on the defect sample information contained in the interaction information, dynamically adjust the confidence threshold or result filtering rules in the post-processing stage of the inference results.

6. The intelligent defect identification and correction method based on interactive target detection according to claim 1, characterized in that: The video structured data in step S4 includes at least one or more of the following: process number, process time information, location, detection result, defect type, and image or video index information.

7. The intelligent defect identification and correction method based on interactive target detection according to claim 1, characterized in that: The actuator mentioned in step S5 includes a manufacturing device, a correction device, or an automated control unit that is communicatively connected to it.

8. The intelligent defect identification and correction method based on interactive target detection according to claim 7, characterized in that: The interactive information is input through a human-computer interaction interface or a host system, and affects the inference results of the target detection model without retraining the complete model.

9. The intelligent defect identification and correction method based on interactive target detection according to claim 1, characterized in that: The inference results of the target detection model include one or more of the following: Process completed, process not completed, no object detected, object in abnormal position.

10. An intelligent defect identification and correction system based on interactive target detection, wherein the system is used based on the intelligent defect identification and correction method based on interactive target detection as described in any one of claims 1-9, characterized in that, include: An industrial camera module, a processor, and a memory, wherein the memory stores a computer program that, when executed on the processor, causes the processor to perform the defect identification method as described in any one of claims 1 to 9.

Images