An intelligent cloth inspection method and system for defect autonomous learning

Abstract

The application relates to the technical field of textile fabric detection, and particularly discloses an intelligent cloth inspecting method and system with self-learning defects, which comprises the following steps: acquiring a current cloth image sequence to be detected; detecting the cloth image sequence to be detected through a preset defect detection model to obtain a detection result set; extracting markable defect targets from the detection result set; and iteratively updating the preset defect detection model based on the markable defect targets. The application detects defects in cloth images, iteratively updates the defect detection model according to the detection results, continuously optimizes the fitting capability of the detection model for various defects, and continuously improves the coverage and recognition accuracy of the defect detection model in the actual application process in a self-learning manner.

Description

Technical Field

[0001] This application relates to the field of textile fabric inspection technology, and in particular to an intelligent fabric inspection method and system with self-learning defect detection. Background Technology

[0002] Textile fabrics undergo quality inspection at every stage of production, primarily detecting defects that occur during weaving and dyeing. Currently, most fabric inspection relies on human observation, which is highly subjective and suffers from low efficiency and high costs. Therefore, with the development of computer vision technology, vision-based fabric defect detection systems will gradually become a crucial solution replacing manual quality inspection.

[0003] Due to the variety of fabric types and color combinations, there are many types of fabric defects. This means that existing deep learning-based defect identification methods require a large amount of sample data to learn the best identification model. However, in actual production environments, the probability of defect samples appearing is low, and it is not practical to sacrifice production efficiency to collect sample data. This makes it difficult for existing defect identification methods to effectively collect enough sample data to learn an identification model that can cover more types. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent fabric inspection method and system for autonomous learning of defects. Through semi-supervised learning, it can continuously learn various types of defects using a small amount of sample data, thereby continuously improving the coverage and recognition accuracy of the defect detection model in actual use.

[0005] Firstly, this application provides an intelligent fabric inspection method based on defect self-learning, employing the following technical solution:

[0006] Obtain the current image sequence of the fabric to be detected;

[0007] The sequence of fabric images to be inspected is detected using a preset defect detection model to obtain a set of detection results;

[0008] Extract tagged defect targets from the detection result set;

[0009] The pre-defined defect detection model is iteratively updated based on the tagged defect target.

[0010] The above technical solution allows for iterative updates to the defect detection model using the defect detection results of fabric images. This continuously optimizes the model's ability to fit various defects and, through self-learning, enhances the coverage and recognition accuracy of the defect detection model during application.

[0011] Optionally, acquiring the image of the fabric to be detected includes:

[0012] Acquire the currently digitized scanned image of the fabric;

[0013] The fabric image is divided into several consecutive and intersecting images, and the divided images are numbered.

[0014] The images to be detected are preprocessed sequentially according to their numbers to form a sequence of images to be detected.

[0015] Optionally, before acquiring the current image of the fabric to be detected, the method further includes:

[0016] Obtain the information of the fabric to be detected, including the fabric type;

[0017] Based on the current fabric type, the types of defects that can be detected are identified using a pre-set textile fabric defect reference table.

[0018] Optionally, the detection results include defect location information, defect category, and confidence level. After the sequence of fabric images to be detected is processed using a preset defect detection model to obtain the detection result set, it includes:

[0019] All detection results in the detection result set are sorted according to the detection image sequence;

[0020] Based on the sorted detection results, the system sequentially determines whether there is any overlap in the detection areas according to the defect location information in the detection results.

[0021] If so, extract a new detection image based on the overlapping defect detection areas;

[0022] The new image is detected using a pre-defined defect detection model, the detection results are obtained, and the detection results are added to the detection result set.

[0023] Optionally, extracting the tagged defect targets from the detection result set includes:

[0024] Determine whether the set of detection results is not empty and whether there are detection results with a confidence level greater than a preset first threshold;

[0025] If so, the detection results in the detection result set with a confidence level greater than the preset first threshold are recorded as candidate targets;

[0026] Further determine whether the defect category of the candidate target belongs to the detectable defect category;

[0027] If so, then the candidate target is marked as a tagged defect target.

[0028] Optionally, the iterative update of the preset defect detection model based on the tagged defect target includes:

[0029] The detection results of the learning process are converted into labeled images based on the defect location information and defect category;

[0030] The labeled images are preprocessed, and the preprocessed images are added to a pre-defined training dataset.

[0031] The pre-defined defect detection model is iteratively trained based on the new training dataset to obtain an updated defect detection model.

[0032] Optionally, after obtaining the detection result set by detecting the sequence of fabric images to be inspected using a preset defect detection model, the method further includes:

[0033] Based on the set of test results, the test results under each defect category are statistically analyzed from the perspective of defect category;

[0034] Based on the defect location information in the inspection results, obtain the size of the defect;

[0035] Based on the existing defect categories, the number of defects under each defect category, and the defect size, an initial score is generated according to preset scoring rules;

[0036] Output the initial score information and associate the detection results with the current fabric information and store them in the preset database.

[0037] Optionally, after storing the detection results in a preset database in association with the current fabric information, the method further includes:

[0038] Determine whether the detection result data corresponding to the current fabric information has reached the preset quantity.

[0039] If so, then the defect types in all inspection results are statistically analyzed, and a statistical data set is generated;

[0040] The causes of defects are analyzed and evaluated based on the statistical data set, and corresponding prompts are generated.

[0041] Secondly, this application provides an intelligent fabric inspection system with self-learning defect detection, comprising:

[0042] Image sequence acquisition module 101 is used to acquire the current image sequence of the fabric to be detected;

[0043] The defect detection module 102 is used to detect the image sequence of the fabric to be inspected using a preset defect detection model and obtain a set of detection results.

[0044] The detection model update module 103 is used to extract tagged defect targets from the detection result set and iteratively update the preset defect detection model based on the tagged defect targets.

[0045] Thirdly, this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above for an intelligent inspection method for autonomous learning of defects.

[0046] In summary, this application first detects defects in fabric images, then filters the detection results to use as new label data for iterative updates to the defect detection model. This optimizes the model's ability to fit defects and allows for continuous improvement in coverage and recognition accuracy through self-learning during practical applications. Furthermore, statistical analysis of the detection results, combined with the detected fabric, allows for scoring to aid in quality inspection. Additionally, the accumulated detection data helps assess the causes of defects, enabling targeted adjustments and optimizations. Attached Figure Description

[0047] Figure 1 This is a flowchart of an intelligent fabric inspection method for defect self-learning provided in an embodiment of this application;

[0048] Figure 2 This is a flowchart illustrating the process of acquiring an image of the fabric to be detected, as provided in an embodiment of this application.

[0049] Figure 3 This is a flowchart of the process for identifying and processing cross-image defects provided in an embodiment of this application;

[0050] Figure 4 This is a flowchart illustrating the extraction of tagged defect targets from a set of detection results, provided in an embodiment of this application.

[0051] Figure 5 This is a flowchart illustrating the iterative update of a preset defect detection model based on a tagged defect target, as provided in an embodiment of this application.

[0052] Figure 6 This is a flowchart provided in an embodiment of the present application for scoring the tested fabric based on the set of test results;

[0053] Figure 7 This is a schematic diagram of an intelligent fabric inspection system that autonomously learns about defects, provided in an embodiment of this application. Detailed Implementation

[0054] The following is in conjunction with the appendix Figure 1 -Appendix Figure 7 This application will be described in further detail below.

[0055] This application provides an intelligent fabric inspection method that autonomously learns about defects; see [link to relevant documentation]. Figure 1 This includes the following steps:

[0056] S100: Obtain the current image sequence of the fabric to be detected.

[0057] Since this application provides an intelligent fabric inspection method that autonomously learns about defects, it can be applied to corresponding fabric inspection devices or equipment. Such equipment generally has unwinding and rewinding mechanisms. By placing the fabric to be inspected in a designated area, the fabric is unwound and inspected through a fabric transmission mechanism. Unlike manual inspection, machine vision is used for inspection, which has a corresponding image acquisition area. For example, when the fabric is unwound, a digital scan is used to simultaneously acquire the fabric image. Then, the fabric image is inspected for defects using a machine vision algorithm. When a defect is detected, a corresponding prompt message is generated so that staff can confirm and record it.

[0058] Therefore, in the application embodiment, the detection of fabric defects first involves acquiring the current fabric image sequence to be detected.

[0059] Specifically, see Figure 2 To obtain an image of the fabric to be inspected, the specific steps include:

[0060] S110. Obtain the currently digitized scanned image of the fabric.

[0061] S120. Divide the fabric image into several consecutive and intersecting images, and number the divided images.

[0062] S130. The images to be detected are preprocessed sequentially according to their numbers to form a sequence of images to be detected.

[0063] Since the fabric is generally unfolded step by step through the fabric transmission mechanism, the area scanned each time is related to the length of the fabric being transported. Therefore, the fabric to be inspected is usually scanned multiple times to achieve comprehensive inspection of each piece of fabric.

[0064] Therefore, in this embodiment, the first step is to acquire the digitally scanned image of the fabric. Additionally, considering that the size of the scanning area depends on the specific device, the scanned image may require significant scaling before it can be used as input to the algorithm's detection model, potentially causing image distortion and affecting defect detection.

[0065] Therefore, after acquiring the digitally scanned fabric image, it is further divided into several continuous and overlapping images, and these images are numbered. The reason for this continuity is to ensure that all areas of the fabric can be detected. Furthermore, considering...

[0066] There may be some defects at the boundaries of the divided images, so a corresponding compensation amount is set for the divided regions, that is, a certain amount of intersection is set between each region to increase the coverage of various defects.

[0067] In addition, to facilitate the identification of the specific location of defects in the fabric based on the detection results, the fabric images are divided and numbered, and then preprocessed sequentially according to these numbers to form a sequence of images to be detected. The image preprocessing mainly involves cropping and scaling the images to match the image input size of the algorithm's detection model. Additionally, image enhancement processes such as noise reduction and lighting compensation are performed to minimize the impact of digital scanning.

[0068] Considering the different types of defects that may occur in different types of fabrics, some types of defects may only appear in a certain type of fabric. For example, uneven needle spacing (causing straight stripes of varying widths on the fabric surface due to inconsistent needle spacing) and misaligned yarn (causing the yarn loops formed by one yarn in the same quilt to be of a different quality than the other yarns after dyeing) may appear in knitted fabrics; cotton grains (dead cotton fiber clumps in the raw cotton that were not completely removed during the cleaning process) and cottonseed husks (cottonseed husks attached to the fiber roots when separating cotton fibers from cotton seeds during the ginning process, or cotton seeds being crushed and mixed into the raw cotton) may only appear in cotton fabrics.

[0069] In addition to considering the relationship between defects and fabric, it is also necessary to take into account that some common types of defects may show different test results in different fabrics. In this case, it will also have a certain impact on the judgment of defect test results.

[0070] Therefore, in this embodiment of the application, when inspecting the fabric, the fabric type is also given certain attention, so as to assist in the judgment of the test results based on the current fabric type. At the same time, it is also convenient to correlate and statistically analyze the defects that appear in various fabrics, so as to carry out corresponding data analysis and inventory.

[0071] Specifically, before acquiring the current image of the fabric to be detected, the following steps are also included:

[0072] S140. Obtain the information of the fabric to be inspected.

[0073] S150. Based on the current fabric type, the detectable defect categories are confirmed through a preset textile fabric defect reference table.

[0074] The fabric information includes fabric type and specifications. Fabric types are categorized into various types, such as cotton, knitted, and synthetic fibers. Fabric specifications indicate the thickness and dimensions of the current fabric, and this data can be entered by staff. Entering fabric specifications is necessary to accurately determine the number of digital scans performed on the fabric, i.e., the number of subsequent image sequences of the fabric to be inspected. Based on this, the specific location of defects in the original fabric can be quickly identified according to the inspection results, facilitating confirmation and verification by relevant staff.

[0075] The pre-defined textile fabric defect reference table represents a statistical analysis of potential defect types for each type of fabric, based on historical data and relevant reference materials. This data is presented in tabular form and stored in a database. It's worth noting that different users may have varying needs for fabric inspection. For example, different users may require different types of fabrics, and different manufacturing processes can also lead to differences in defect types. Therefore, the defect reference table should be tailored to the specific usage scenario and requirements.

[0076] In this embodiment of the application, when inspecting each piece of fabric, the fabric information is first obtained, and then based on the current fabric type, all defect categories that can be detected for the current fabric type are confirmed by referring to the textile fabric defect reference table stored in the database.

[0077] S200: The sequence of fabric images to be inspected is inspected using a preset defect detection model to obtain a set of inspection results.

[0078] The preset defect detection model is a preliminary detection model trained on a self-made or existing defect image dataset, based on an object detection algorithm. The object detection algorithm can be a classic object detection algorithm such as FAST-RCNN or SSD; this application does not limit its scope.

[0079] The detection result set represents the combination of detection results generated for each image in the sequence of fabric images to be detected. The detection results include defect location information, defect category, and confidence level. Defect location information is equivalent to the predicted region of the target, reflecting the location and size of the defect in the image. The defect category is the type of defect within the predicted target region. The confidence level is equivalent to a probability, indicating the likelihood that the predicted target belongs to a particular defect; a higher confidence level indicates a greater probability that the predicted target belongs to a certain defect.

[0080] Because different fabric types and color combinations result in a wide variety of fabric defects, and the probability of defect samples appearing in actual production environments is generally low, collecting defect sample data is inherently difficult. In addition, the imbalance of various defect sample data leads to a low fitting degree and low coverage of defect types in deep learning-based defect identification methods.

[0081] Therefore, in this embodiment, a self-training approach from semi-supervised learning is used to continuously increase the number of trainable samples, thereby continuously improving the fitting and generalization capabilities of the defect detection model. Specifically, the model is first trained using existing defect sample data (the original training dataset) to obtain an initial defect detection model. Then, the current fabric image sequence to be detected is analyzed, and the detection results are obtained. Finally, new training sample data is injected into the entire detection model based on the detection results, thereby updating and iterating the current detection model.

[0082] Since the fabric image to be inspected is a digitally scanned and segmented image of the fabric, and the size of the defects varies, there may be differences between defects of the same type and different types. Therefore, there may be defects that appear across images, such as broken weft or broken warp. In this case, the inspection results of such defects are prone to deviation.

[0083] Therefore, in order to reduce the impact of this situation on defect detection, after obtaining the detection results, the positional relationship of each detection target area in the detection results will be judged, and a new image segmentation will be given for targets that appear across images, so that the target can be re-detected.

[0084] Specifically, see Figure 3 After detecting the fabric image sequence to be inspected using a preset defect detection model and obtaining the detection result set, the following steps are also included:

[0085] S210. Sort all detection results in the detection result set according to the detection image sequence.

[0086] S220. Based on the sorted detection results, determine whether there is any overlap in the detection areas according to the defect location information in the detection results.

[0087] S230. If so, extract a new detection image based on the overlapping defect detection areas;

[0088] S240. The new detection image is detected using a preset defect detection model, the detection results are obtained, and the detection results are added to the detection result set.

[0089] In this embodiment of the application, after the sequence of fabric images to be inspected is inspected by a preset defect detection model and a set of inspection results is obtained, all the inspection results in the set of inspection results are sorted according to the sequence of inspection images. This is because the defect location information presented in the inspection results will contain the coordinate information of the inspection area. Combined with the division method of the original scanned image, the specific location coordinates of each defect inspection area in the original scanned image can be determined.

[0090] This allows all defect location information in the detection results, i.e., the target detection area, to be mapped onto the original scanned image. Then, it can be determined whether there is overlap in the target detection areas. If there is overlap, a new detection image is extracted based on the overlapping defect detection areas. This means merging the overlapping target detection areas to create a new detection image. Finally, the new detection image is tested using a preset defect detection model to obtain the corresponding detection results, which are then added to the detection result set.

[0091] S300: Extract tagged defect targets from the detection result set.

[0092] Among them, the labelable defect target refers to the image that can be labeled with defects. After appropriate processing, it can be used as sample data for training. The so-called sample data for training is the fabric image after labeling.

[0093] In this embodiment of the application, the aim is to extract new defect sample data through the defect detection results of the actual application process, and then add the new sample data to the original training dataset to iteratively update the current defect detection model.

[0094] Specifically, see Figure 4 Extracting tagged defect targets from the detection result set includes the following steps:

[0095] S310. Determine whether the detection result set is not empty and whether there are detection results with a confidence level greater than the preset first threshold.

[0096] S320. If so, the detection results in the detection result set with a confidence level greater than the preset first threshold are recorded as candidate targets.

[0097] S330, Further determine whether the defect category of the candidate target belongs to the detectable defect category.

[0098] S340. If so, then the candidate target is marked as a markable defect target.

[0099] Since the labelable defect targets will be used in subsequent model updates and training, certain conditions must be met. First, the sample images in the training dataset are usually manually labeled, which is equivalent to establishing a set of rules. Machine learning is then used to apply these rules to the identification and differentiation of defects.

[0100] Therefore, the labels of the most reliable samples should be used as the true labels for the newly added training samples. In other words, the labelable defect target must first have a certain recognition accuracy, that is, the confidence level must be high. This means that all the results detected by the current detection model will be judged. Of course, there must be detection results first. If the current fabric has no defects, it may not be able to detect anything, that is, the final set of detection results will be an empty set.

[0101] Therefore, the system first determines whether the detection result set is not empty and whether there are detection results with a confidence level greater than a preset first threshold. Only detection results with a confidence level greater than the preset first threshold are considered eligible to participate in the update training and are thus recorded as candidate targets. The preset first threshold here represents the standard for measuring the confidence level of the detection results, but this standard is not fixed. As mentioned above, some defects may present different detection results in different fabrics. Therefore, the preset first threshold can also be adjusted according to the type of fabric being detected.

[0102] As mentioned above, considering that different types of fabrics may have different types of defects, and that some types of defects may only appear in a certain type of fabric, the type of fabric to be tested is also obtained, and the types of defects that can be detected are confirmed according to the fabric type through a preset textile fabric defect reference table.

[0103] Therefore, after selecting candidate targets with a confidence level that reaches the preset first threshold, it will further determine whether the defect type corresponding to the candidate target is a detectable defect type. If so, the candidate target will be recorded as a tagged defect target; if not, it means that the currently detected defect target is unlikely to appear in the current fabric and the confidence level is high, indicating that it is a false detection. In addition to removing the corresponding candidate targets, the corresponding detection result information will also be recorded so as to conduct subsequent data analysis and inventory for the falsely detected defect type and fabric type.

[0104] S400: Iteratively update the preset defect detection model based on the tagged defect target.

[0105] Specifically, see Figure 5 The pre-defined defect detection model is iteratively updated based on the tagged defect target, including the following steps:

[0106] S410. The detection results to be learned are converted into label images based on the defect location information and defect category.

[0107] S420. Perform data preprocessing on the labeled images and add the preprocessed images to the preset training dataset.

[0108] S430. Iteratively train the preset defect detection model based on the new training dataset to obtain the updated defect detection model.

[0109] In this embodiment, the preset defect detection model is iteratively updated based on the labelable defect target. The first step is to convert the labelable defect target into trainable data samples. Since the labelable defect target represents a detection region of the fabric image (or a portion of the image), and this detection region also includes its defect category, it is essentially a labeled sample image. By simply using the target detection region as a template to crop from the fabric image, the cropped image can be used as a trainable labeled image.

[0110] Next, the label image is preprocessed. This preprocessing is different from the previous preprocessing of the scanned image. The preprocessing here mainly expands the sample image. For example, the sample image is rotated at different angles and has different brightness to generate a variety of sample data. Correspondingly, the amount of data of each defect type in the original sample data will be balanced and adjusted accordingly.

[0111] Finally, the preprocessed image will be added to the preset training dataset (if it is the first update, it will be the original training dataset). The preset defect detection model (the current defect detection model) will be iteratively trained based on the new training dataset. That is, the current defect detection model is used as the pre-trained model, and iterative training will continue on this basis to obtain the updated defect detection model.

[0112] It's worth noting that the model isn't iteratively updated after each detection. Instead, updates are based on actual usage, and the detection of fabric defects and the iterative training of the detection model are separate and do not interfere with each other. For example, if a preliminary detection model is used during the current working period, new sample data is extracted using the detection results obtained during that period after the working period ends. At this point, it's necessary to determine whether the acquired sample data has reached a certain amount. If it has, the detection model will be iteratively trained to obtain an updated detection model. If it hasn't reached a certain amount, data accumulation will continue, because model training also consumes corresponding human and material resources. If the amount of newly added sample data is too small, the results will be relatively low.

[0113] After obtaining the updated detection model, replace the previous detection model with the updated model without affecting its actual operation. The work mentioned here mainly refers to the operation of the fabric inspection equipment.

[0114] Machine vision can identify the presence and type of defects in fabrics, reducing the burden of manual inspection. However, as a fabric inspection system, each defect still needs to be scored to determine whether the inspected fabric passes inspection. In this aspect, machine vision serves primarily as an aid; the final quality assessment still requires human evaluation. Therefore, machine vision's role is to provide more accurate inspection information and perform multi-dimensional analysis of various inspection data to assist relevant personnel in evaluating fabric quality.

[0115] Therefore, after detecting the fabric image sequence using a preset defect detection model and obtaining the detection result set, see [link to relevant documentation]. Figure 6 It also includes the following steps:

[0116] S510. Based on the set of test results, statistically analyze the test results for each defect category from the perspective of defect category.

[0117] S520. Based on the defect location information in the inspection results, obtain the size of the defect.

[0118] S530. Based on the existing defect categories, the number of defects under each defect category, and the defect size, generate an initial score according to the preset scoring rules.

[0119] S540. Output the initial scoring information and associate the detection results with the current fabric information and store them in the preset database.

[0120] The preset scoring rules represent the inspection standards for textile fabrics. For example, how many points are deducted for each occurrence of a certain type of defect, and how points are deducted for certain types of defects based on size. For example, one point is deducted for defects that are no longer than 3 inches, and two points are deducted for defects that are longer than 3 inches but no longer than 6 inches.

[0121] Because the final inspection results presented during the inspection of defect targets are filtered by confidence level, the final results all indicate the presence of defects and can be used as a comprehensive quality assessment of the fabric. However, higher requirements are placed on the selection of tagged defect targets.

[0122] Therefore, after obtaining the set of detection results, the detection results will be statistically analyzed according to the defect category dimension. The detection results under each defect category will be counted. The number of defects can be obtained based on the number of results. Based on the defect location information in the detection results, combined with the scaling ratio of the original scanned image and the image used as input to the detection model, the size of the defect can be obtained.

[0123] Based on the defect categories, the number of defects under each defect category, and the defect size in the current test result set, an initial score can be generated according to the preset scoring rules. The initial score information will be output so that relevant staff can be informed and make appropriate decisions.

[0124] After all the initial scores of a piece of fabric have been tested, a final score will be given based on the information collected during the initial testing. The total score will also be output as a reference for the relevant staff to make a comprehensive evaluation of the fabric.

[0125] In addition, the test results will be linked to the current fabric information and stored in a pre-set database for subsequent data statistical analysis.

[0126] Since different processes or procedures may lead to different types of defects, simply inspecting the finished fabric is not enough to reflect which process caused the defect.

[0127] Therefore, in this embodiment of the application, the results of each test will also be stored so as to help infer the cause of the defect through data analysis.

[0128] Specifically, after storing the test results in a pre-set database along with the current fabric information, the following steps are also included:

[0129] S610. Determine whether the detection result data corresponding to the current fabric information has reached the preset quantity.

[0130] S620. If so, then perform statistics on the defect types in all inspection results and generate a statistical data set.

[0131] S630. Analyze and evaluate the causes of defects based on the statistical data set, and generate corresponding prompt information.

[0132] Since we want to infer the causes of defects based on data statistics, we need a sufficient amount of data to support this. This means that we need to accumulate a certain amount of inspection results for various types of fabrics. When the number of test results for a certain type of fabric reaches a preset amount, we can statistically analyze the types of defects in all test results and generate a statistical data set. Then, we can analyze and evaluate the causes of defects based on the statistical data set and generate corresponding prompts.

[0133] Using a statistical dataset, all types of defects produced by this type of fabric can be clearly reflected, along with the frequency of each defect type. For defect types that occur significantly more frequently than other types, an inventory can be conducted based on the fabric's production process and the defects that are prone to occur in each process. This can roughly deduce the reasons for the high frequency of defect types. After outputting the corresponding evaluation information as prompts, it can help relevant personnel check the associated production processes. If problems occur in the process steps, adjustments can be made in a timely manner.

[0134] This application also provides an intelligent fabric inspection system that autonomously learns about defects, see [link to relevant documentation]. Figure 7 The system includes: an image sequence acquisition module 101, a defect detection module 102, and a detection model update module 103.

[0135] The image sequence acquisition module 101 is used to acquire the current image sequence of the fabric to be detected;

[0136] The defect detection module 102 is used to detect the image sequence of the fabric to be inspected using a preset defect detection model and obtain a set of detection results.

[0137] The detection model update module 103 is used to extract tagged defect targets from the detection result set and iteratively update the preset defect detection model based on the tagged defect targets.

[0138] In this embodiment of the application, the image sequence acquisition module 101 is specifically used to acquire the original image of the fabric to be inspected by digital scanning, and to divide the original image to obtain the current image sequence of the fabric to be inspected.

[0139] The defect detection module 102 is specifically used to detect the image sequence of the fabric to be inspected using a preset defect detection model and obtain a set of detection results.

[0140] The detection model update module 103 is specifically used to extract labelable defect targets from the detection result set, convert the labelable defect targets into trainable defect sample images, and then iteratively update the preset defect detection model by expanding the original sample training dataset.

[0141] This application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as any of the above-described intelligent fabric inspection methods for autonomous learning of defects.

[0142] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the principles of this application should be covered within the scope of protection of this application.

Claims

1. A method for intelligent fabric inspection that autonomously learns about defects, characterized in that, include: Obtain the current image sequence of the fabric to be detected; The sequence of fabric images to be inspected is detected using a preset defect detection model to obtain a set of detection results; Extract tagged defect targets from the detection result set; The pre-set defect detection model is iteratively updated based on the tagged defect target; The detection results include defect location information, defect category, and confidence level. The sequence of fabric images to be detected is processed using a preset defect detection model, and the resulting set of detection results includes: All detection results in the detection result set are sorted according to the image sequence of the fabric to be detected; Based on the sorted detection results, the system sequentially determines whether there is any overlap in the detection areas according to the defect location information in the detection results. If so, extract a new image of the fabric to be inspected based on the overlapping defect detection areas; The new fabric image to be inspected is detected using a preset defect detection model, the detection results are obtained, and the detection results are added to the detection result set. Based on the set of test results, the test results under each defect category are statistically analyzed from the perspective of defect category; Based on the defect location information in the inspection results, obtain the size of the defect; Based on the existing defect categories, the number of defects under each defect category, and the defect size, an initial score is generated according to preset scoring rules; Output initial scoring information and associate the detection results with the current fabric information into a preset database; The extraction of tagged defect targets from the detection result set includes: Determine whether the set of detection results is not empty and whether there are detection results with a confidence level greater than a preset first threshold; If so, the detection results in the detection result set with a confidence level greater than the preset first threshold are recorded as candidate targets; Further determine whether the defect category of the candidate target belongs to the detectable defect category; If so, then mark the candidate target as a tagged defect target; After storing the detection results in a preset database in association with the current fabric information, the method further includes: Determine whether the detection result data corresponding to the current fabric information has reached the preset quantity. If so, then the defect types in all inspection results are statistically analyzed, and a statistical data set is generated; The causes of defects are analyzed and evaluated based on the statistical data set, and corresponding prompts are generated. The step of obtaining the current fabric image sequence to be detected includes: Acquire the currently digitized scanned image of the fabric; The fabric image is divided into several consecutive and intersecting images, and the divided images are numbered. The images are preprocessed sequentially according to their numbers to form a sequence of fabric images to be detected. Specifically, all detection results in the detection result set are sorted according to the sequence of fabric images to be detected. Based on the sorted detection results, the system sequentially determines whether there is overlapping detection areas based on the defect location information in the detection results. If so, a new fabric image to be detected is extracted based on the overlapping defect detection areas, including: For all detection results in the detection result set, sort them according to the sequence of fabric images to be detected, and then determine the specific location coordinates of each defect detection area in the fabric image based on the method of dividing the fabric image. Map all defect location information in the detection results, i.e., the target detection area, to the fabric image and determine whether there is any overlap of target detection areas. If there is overlap of target detection areas, extract a new fabric image to be detected based on the overlapping defect detection areas. That is, merge the fabric images to be detected with overlapping target detection areas as a new fabric image to be detected.

2. The intelligent fabric inspection method based on defect self-learning according to claim 1, characterized in that, Before acquiring the current fabric image sequence to be detected, the method further includes: Obtain the information of the fabric to be detected, including the fabric type; Based on the current fabric type, the types of defects that can be detected are identified using a pre-set textile fabric defect reference table.

3. The intelligent fabric inspection method based on defect self-learning according to claim 1, characterized in that, The iterative update of the preset defect detection model based on tagged defect targets includes: The detection results to be learned are converted into labeled images based on the defect location information and defect category; The labeled images are preprocessed, and the preprocessed images are added to a pre-defined training dataset to obtain a new training dataset. The pre-defined defect detection model is iteratively trained based on the new training dataset to obtain an updated defect detection model.

4. A defect-autonomous learning intelligent fabric inspection system, applied to the defect-autonomous learning intelligent fabric inspection method as described in any one of claims 1 to 3, characterized in that, include: The image sequence acquisition module (101) is used to acquire the current image sequence of the fabric to be detected; The defect detection module (102) is used to detect the image sequence of the fabric to be inspected using a preset defect detection model and obtain a set of detection results. The detection model update module (103) is used to extract the markable defect targets from the detection result set and iteratively update the preset defect detection model based on the markable defect targets.

5. A computer-readable storage medium storing a computer program that is loaded by a processor and executes a smart inspection method for defect self-learning as described in any one of claims 1 to 3.

Images