Cone yarn direction detection method and system based on image visual identification technology

Abstract

The invention belongs to the technical field of computer vision, and provides a cone yarn direction detection method and system based on an image visual identification technology. The method comprises the following steps: pre-constructing a special model library comprising a plurality of target detection models; extracting image features and confidence from the real-time image of the cone yarn, and performing enhancement transformation processing on the image features according to the confidence to obtain a plurality of enhanced image features; a plurality of related target detection models are matched from the special model library according to the enhanced image features, and a dynamic composite model is established based on the target detection models; and processing the real-time image by using the dynamic composite model to obtain a predicted bounding box of the yarn head area, and judging the direction of the cone yarn according to the position of the predicted bounding box in the real-time image. According to the invention, accurate positioning and direction determination of the yarn head area can be realized in a changeable production environment.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to a method and system for detecting yarn bobbin orientation based on image visual recognition technology. Background Technology

[0002] In the textile industry, yarn packages are a widely used form of yarn packaging, consisting of cylindrical or conical yarns wound onto bobbins according to specific patterns. As a core raw material in weaving preparation processes (such as warping and winding), the level of automation and intelligence in the processing of yarn packages directly determines the production efficiency and product quality of the entire textile factory.

[0003] In automated warping processes, the automatic loading (coiling) and unloading of yarn cones is a crucial step. This process relies on automated guided vehicles (AGVs) or robotic arms to accurately pick up the yarn cones and place them in designated positions on the warping frame. However, yarn cones have a clear directionality: one end is a fixed bobbin base, and the other end is a pull-out yarn end (referred to as "yarn end"). Correctly identifying the direction of the yarn cone is a prerequisite for ensuring that the yarn can be pulled out smoothly and without tangling. If the direction is incorrectly identified, it will lead to a series of serious problems, including but not limited to: (1) the yarn cannot be pulled out or breaks, causing production interruption; (2) uneven warp tension, resulting in fabric defects and affecting fabric quality; (3) automated equipment (such as AGVs) performs incorrect picking and placing actions, requiring manual intervention for correction, thus making it impossible to achieve true "lights-out factory" full-process automation.

[0004] However, achieving automated identification in real-world industrial scenarios faces numerous technical challenges, with the core challenge concentrated on the visual recognition of yarn ends. Specifically, during production and handling, yarn ends may be pressed into the yarn layer, hang naturally, or attach to the surface of the yarn package in any arbitrary form, lacking fixed and uniform morphological features. This makes it difficult for traditional rule-based image processing algorithms to stably locate and extract them. Furthermore, changes in lighting and complex background noise generated by the yarn texture on the yarn package surface can severely interfere with the accurate identification of this small target. Conventional machine vision methods (such as those based on edge detection and template matching) exhibit poor robustness, high false positive and false negative rates when faced with the aforementioned weak and variable target features and complex imaging environments, making it difficult to meet the high-speed and high-reliability production requirements of modern intelligent textile factories.

[0005] Therefore, there is an urgent need for a solution that can overcome the above difficulties, adapt to complex industrial environments, and achieve rapid, accurate, and stable identification of the yarn bobbin direction. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for detecting the yarn bobbin direction based on image visual recognition technology.

[0007] This invention provides a method for detecting the direction of yarn bobs based on image visual recognition technology, comprising the following steps: First, a dedicated model library comprising multiple target detection models is pre-constructed, each target detection model being trained using a corresponding set of dedicated image data; wherein the morphological features of the yarn end and the background features of the yarn bobbin corresponding to each dedicated image data are similar; second, a real-time image of the yarn bobbin to be tested is acquired, and image features and their confidence scores are extracted from the real-time image, the image features including yarn end morphological features and yarn bobbin background features; third, the image features are enhanced and transformed according to the confidence scores to obtain several enhanced image features; fourth, multiple related target detection models are matched from the dedicated model library based on the enhanced image features, and a dynamic composite model is constructed based on each target detection model; fifth, the dynamic composite model is used to process the real-time image to obtain a predicted bounding box of the yarn end region, and the direction of the yarn bobbin is determined based on the position of the predicted bounding box in the real-time image.

[0008] This invention also provides a yarn bobbin orientation detection system based on image visual recognition technology. The system includes: a model building unit, which pre-builds a dedicated model library including multiple target detection models, each target detection model being trained using a corresponding set of dedicated image data; wherein the yarn end morphological features and yarn bobbin background features corresponding to each dedicated image data are similar; a feature enhancement processing unit, which acquires a real-time image of the yarn bobbin to be tested, extracts image features and their confidence scores from the real-time image, the image features including yarn end morphological features and yarn bobbin background features; and performs enhancement transformation processing on the image features according to the confidence scores to obtain several enhanced image features; a model matching and synthesis unit, which matches multiple related target detection models from the dedicated model library according to the enhanced image features, and constructs a dynamic composite model based on each target detection model; and a recognition and inference unit, which processes the real-time image using the dynamic composite model to obtain a predicted bounding box of the yarn end region, and determines the orientation of the yarn bobbin based on the position of the predicted bounding box in the real-time image.

[0009] The beneficial technical effects of this invention are at least as follows: By pre-constructing a dedicated model library for different yarn end shapes and background features, this invention effectively solves the problem of insufficient generalization ability of a single model in complex industrial scenarios; and, through confidence-based adaptive feature enhancement technology, it significantly improves the representation ability of low-quality image features and enhances the robustness of model matching; and, by adopting a dynamic composite model mechanism, it can intelligently combine the optimal detection model according to the features of the image to be tested, achieving a significant improvement in detection accuracy and adaptability. Ultimately, it achieves accurate positioning and orientation determination of the yarn end area in a variable production environment, providing reliable technical support for subsequent automated operations and greatly improving the intelligence level and operational efficiency of the textile production process. Attached Figure Description

[0010] Figure 1 This is a flowchart of a yarn bobbin orientation detection method based on image visual recognition technology disclosed in an embodiment of the present invention.

[0011] Figure 2 This is a structural diagram of a yarn bobbin orientation detection system based on image visual recognition technology disclosed in an embodiment of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0014] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0015] like Figure 1 As shown, this embodiment of the invention discloses a yarn bobbin direction detection method 100 based on image visual recognition technology, including the following steps: 10, a dedicated model library including multiple target detection models is pre-constructed, and each target detection model is trained using a corresponding set of dedicated image data; wherein, the yarn end morphological features and yarn bobbin background features corresponding to each dedicated image data are similar.

[0016] In this step, to address the issue of insufficient generalization ability of a single model when faced with diverse yarn end shapes and complex backgrounds, this invention constructs a specialized model library with strong targeting and broad coverage. Specifically, the specialized model library is constructed as follows: a large number of yarn package images are collected and finely classified based on yarn end morphological features (e.g., yarn ends pressed into the yarn layer, hanging naturally, or attached to the yarn package surface in any shape) and yarn package background features (e.g., different lighting conditions, yarn package surface texture noise). Images with similar dominant features (i.e., yarn end morphological features and yarn package background features) are grouped together to form multiple sets of specialized image data, which serve as training datasets. Each set of specialized image data is used to train a separate object detection model (such as the YOLO series models). Finally, these object detection models trained for different specific working conditions together form the specialized model library, with each object detection model excelling at identifying yarn ends in a specific type of image.

[0017] 20. Acquire a real-time image of the yarn package to be tested, extract image features and their confidence levels from the real-time image, the image features including yarn end morphological features and yarn package background features; and perform enhancement transformation processing on the image features according to the confidence levels to obtain several enhanced image features.

[0018] In this step, real-time images of the yarn packages to be inspected on the production line are first acquired, and key image features (such as the possible shape and color of the yarn ends, as well as the texture and lighting conditions of the background) and their corresponding confidence scores are extracted. The confidence score is used to quantify the salience and reliability of the feature in the image.

[0019] Next, based on the confidence level, the extracted original image features undergo differentiated enhancement transformations. For example, for image features with high confidence, a light enhancement transformation is performed to fine-tune and explore within a small range while maintaining their reliable information, preventing mismatch with the model due to minor fluctuations. For image features with low confidence, a deep enhancement transformation is performed to generate multiple differentiated feature representations through large-scale, multi-directional transformations, thereby compensating for the uncertainty of the original features and increasing their probability of matching the correct dedicated model.

[0020] After the confidence-based enhancement transformation process described above, multiple enhanced image features are derived from one image feature, serving as multiple possible representations of the real-time image.

[0021] 30. Based on the enhanced image features, multiple relevant target detection models are matched from the dedicated model library, and a dynamic composite model is constructed based on each target detection model.

[0022] In this step, the enhanced image features obtained in step 20 are compared and similarity is calculated with the feature representations (i.e., yarn morphology features and yarn background features) corresponding to each target detection model in the dedicated model library. Target detection models that are frequently selected with high similarity by multiple enhanced feature versions are filtered out.

[0023] Subsequently, these selected expert models are combined to form a dynamic composite model. This composite model is customized for the specific characteristics of the current image to be tested (reflected by its enhanced image features), and its comprehensive detection capability is superior to any single fixed model. Understandably, the prediction results output by each object detection model in this dynamic composite model are integrated through an internal fusion mechanism to obtain the final predicted bounding box.

[0024] 40. The dynamic composite model is used to process the real-time image to obtain the predicted bounding box of the yarn end region. The direction of the yarn package is determined based on the position of the predicted bounding box in the real-time image.

[0025] In this step, the real-time image of the yarn to be tested is input into the dynamically constructed dynamic composite model in step 30. This dynamic composite model will coordinate the detection results of its internal expert models and perform comprehensive reasoning through internal fusion mechanisms (such as weighted voting, non-maximum suppression, etc.) to finally output the most reliable predicted bounding box for the yarn end region.

[0026] Based on the specific location of the predicted bounding box in the real-time image (e.g., on the left or right side, top or bottom of the yarn bobbin), and combined with the pre-set rules, the correct direction of the yarn bobbin is finally determined, providing accurate guidance information for the subsequent automatic grasping and yarn loading operations of the AGV or robotic arm.

[0027] This invention effectively addresses the problem of insufficient generalization ability of a single model in complex industrial scenarios by pre-constructing a dedicated model library for different yarn end shapes and background features. Furthermore, it significantly improves the representation ability of low-quality image features and enhances the robustness of model matching through confidence-based adaptive feature enhancement technology. Finally, it employs a dynamic composite model mechanism to intelligently combine the optimal detection model based on the features of the image under test, achieving a significant improvement in detection accuracy and adaptability. Ultimately, this invention enables precise positioning and orientation determination of yarn end areas in variable production environments, providing reliable technical support for subsequent automated operations and greatly improving the intelligence level and operational efficiency of the textile production process.

[0028] As an example, each object detection model is trained using a corresponding set of dedicated image data, including: 11, collecting yarn package images that meet the quantity conditions, and classifying all yarn package images according to the morphological features of the yarn ends and the background features of the yarn packages; wherein, the morphological features of the yarn ends include the yarn ends being pressed into the yarn layer, hanging naturally, or attached to the surface of the yarn packages in any shape; the background features of the yarn packages include different lighting conditions and the texture of the yarn package surface.

[0029] This step requires collecting a massive number of yarn bobbin images, and these images must meet the quantity requirement, that is, ensure that the image data used to train each type of specialized model is sufficient in quantity, so as to avoid problems such as model overfitting or weak generalization ability due to insufficient data.

[0030] All yarn package images are systematically classified based on the morphological characteristics of the yarn ends and the background characteristics of the yarn packages. The morphological characteristics of the yarn ends mainly involve the physical state of the yarn ends themselves, including typical cases such as the yarn ends being pressed into the yarn layer, hanging naturally, or attached to the surface of the yarn packages in any shape; while the background characteristics of the yarn packages mainly involve the influence of the imaging environment, including different lighting conditions (such as strong light, weak light, shadows and reflections caused by side lighting) and the complex texture of the yarn package surface itself.

[0031] 12. Group yarn package images with similar yarn end morphological features and similar yarn package background features into a group to form a set of dedicated image data. Use this set of dedicated image data to train a target detection model.

[0032] In this step, based on the completed feature classification, the yarn package images are grouped to form the final training set. Specifically, the images classified in step 11 are grouped according to the principle of similar yarn end morphological features and similar yarn package background features. For example, all images showing the natural draping of the yarn end against a strong light background are grouped into one group; all images showing the yarn end being pressed into the yarn layer against a complex surface texture background are grouped into another group.

[0033] Therefore, each set of dedicated image data possesses a high degree of feature consistency and scene specificity. Using such a highly homogeneous set of data to train a specific object detection model allows the model to focus on learning and recognizing yarn patterns under specific working conditions, thus becoming an expert model for solving such specific recognition problems. This significantly improves the detection accuracy and robustness of object detection models in their respective areas of expertise.

[0034] In this embodiment, a systematic classification and grouping of massive images based on yarn end morphological features and yarn bobbin background features is performed beforehand to construct multiple dedicated image datasets with highly consistent internal features. This method ensures that each target detection model is trained using homogeneous data specific to a particular working condition, enabling the model to focus highly on learning and accurately identifying yarn end patterns in that specific scenario. This refined data construction strategy fundamentally improves the detection accuracy and robustness of each dedicated model in its area of ​​expertise, thereby helping subsequent dynamic composite models achieve high-precision and highly adaptable yarn bobbin direction detection, effectively overcoming the deficiency of insufficient model generalization ability caused by mixed training data.

[0035] As an example, extracting the confidence level of the image features from the real-time image includes: 21, evaluating the saliency of the image features based on the degree of clustering of the image features in the feature space, the degree of clustering being determined by calculating the distance between the image features and the cluster centers of similar features; and evaluating the reliability of the image features based on the stability of the image features in adversarial verification, the stability being determined by calculating the degree of change of the image features after adding a small perturbation to the real-time image.

[0036] In this step, this implementation constructs a two-dimensional feature quality evaluation system, which refines the confidence of image features from two aspects: intrinsic nature and external anti-interference capability. Specifically, in terms of saliency evaluation: the degree of clustering is quantified by calculating the Euclidean or cosine distance between the current image feature vector and the pre-established cluster centers of similar features in the feature space. The smaller the distance, the closer the current feature is to the core region of its category in the feature space, and the more typical and prominent its feature expression is, i.e., the higher its saliency; conversely, the larger the distance, the more likely the feature is to be a marginal case or an outlier, and its saliency is lower. It can be understood that this evaluation reflects the typicality or purity of the image feature as a representative of its category.

[0037] In terms of reliability assessment: the stability of the features is tested through adversarial verification. Specifically, a series of imperceptible micro-perturbations (such as Gaussian noise) are added to the original real-time image, then the image features are re-extracted, and the degree of change between the feature vectors before and after the perturbations is calculated (e.g., calculating the L2 norm distance). If the degree of change is small, it indicates that the feature is insensitive to perturbations and has strong robustness, i.e., high reliability; if the degree of change is large, it indicates that the feature is fragile, unstable, and has low reliability. Understandably, this assessment characterizes the robustness of the image feature in maintaining consistency in real-world, complex industrial environments.

[0038] 22. The confidence level of the image feature is calculated based on a weighted combination of the saliency and the reliability.

[0039] In this step, the evaluation results from the two dimensions mentioned above are fused to obtain a comprehensive confidence score. Specifically, appropriate weights are assigned to significance and reliability, for example, depending on the needs of the specific application scenario, the weight of significance can be set to 0.6 and the weight of reliability to 0.4, and the final confidence score is calculated by weighted summation.

[0040] Confidence level = (significance score × Ws) + (reliability score × Wr), where Ws + Wr = 1.

[0041] This weighted combination mechanism ensures that the confidence level not only reflects the clarity (i.e., significance) of the feature itself, but also takes into account its ability to resist interference in practical applications (i.e., reliability).

[0042] This implementation method achieves a refined measurement of image feature confidence by constructing a two-dimensional evaluation system based on saliency and reliability. The evaluation mechanism adopted in this implementation method fully considers the intrinsic properties and external anti-interference capabilities of image features, making the subsequent confidence-based adaptive image feature enhancement transformation more accurate, thereby helping to accurately identify the yarn bobbin direction.

[0043] As an example, the reliability of the image features in adversarial verification is evaluated based on their stability, including: 211, determining the intensity and number of small perturbations based on the image quality score of the real-time image and the neighborhood density of the image features in the feature space.

[0044] This step introduces a two-factor adaptive perturbation parameter determination mechanism. Specifically, it considers the following two key factors simultaneously: the image quality score of the real-time image, which is calculated based on a comprehensive evaluation of the image's sharpness, contrast, noise level, and other indicators. A low image quality score indicates that the original image itself may have problems such as blurriness or excessive noise, making the stability of its features questionable. Therefore, stronger and more numerous perturbations are required for more rigorous stress testing.

[0045] Meanwhile, the neighborhood density of image features in the feature space is introduced as another key factor. Neighborhood density is quantified by counting the number of other training feature points within a given radius centered on the feature point. Specifically, when the neighborhood density is low, it indicates that the feature is located in a relatively sparse or remote region in the feature space, which may be an undertrained or rare feature pattern. Its stability is also questionable, so stronger and more frequent perturbations are needed to fully test whether its foothold in the feature space is secure.

[0046] By comprehensively considering image quality (reflecting uncertainty at the data level) and feature neighborhood density (reflecting cognitive uncertainty at the model level), the intensity and quantity of perturbations required for adversarial verification can be determined more comprehensively and accurately in a dynamic manner.

[0047] 212. The determined micro-perturbation is added to the real-time image to generate a perturbed image. The perturbed image features are extracted again for each perturbed image.

[0048] In this step, based on the perturbation intensity and number determined in step 211, a series of small perturbations that meet the requirements are generated (e.g., Gaussian noise of different amplitudes, salt-and-pepper noise, or small affine transformations, etc.). These perturbations are then superimposed one by one onto the original real-time image to generate several perturbed images. Understandably, although these perturbed images are visually only slightly different from the original image, they are sufficient to pose a challenge to the feature extraction process. Using the exact same feature extraction method as for the original image, feature extraction is performed again on each perturbed image to obtain the corresponding perturbed image features.

[0049] This implementation can be understood as simulating minute changes in images caused by factors such as changes in lighting and equipment vibration in a real industrial environment, and observing the output performance of the feature extraction algorithm under these changes.

[0050] 213. The reliability of the image features is evaluated by calculating the degree of difference between each of the perturbed image features and the original image features.

[0051] In this step, all the perturbed image features obtained in step 212 are compared with the original image features to calculate the degree of difference between them. Specifically, the Euclidean or cosine distance between each perturbed feature and the original feature in the feature space can be calculated, and then these distance values ​​are statistically analyzed (e.g., the average or maximum value is calculated). If the degree of difference is generally small, it indicates that even with minor perturbations in the image, the extracted features remain highly consistent, and the reliability of the original image feature is high; if the degree of difference is large, it indicates that the feature is sensitive to perturbations, has poor stability, and low reliability.

[0052] Finally, based on the statistical results of the above differences, a quantified reliability score is output as input for subsequent comprehensive confidence calculation.

[0053] As an example, the enhancement transformation processing of the image features based on the confidence level to obtain several enhanced image features includes: 23, applying a first type of enhancement transformation to the first type of image features with a confidence level higher than a preset high threshold; the first type of enhancement transformation is a random perturbation within a first intensity range applied to the image features in the feature space.

[0054] This step performs fine-tuning in the feature space for high-confidence Class 1 image features.

[0055] Specifically, when the confidence level of the extracted image features (i.e., yarn morphology features and bobbin background features, typically represented as high-dimensional feature vectors) is higher than a preset high threshold, it indicates that the position of the feature vector in the feature space is reliable and accurate. At this point, the applied first-type enhancement transformation is defined as: superimposing a random perturbation (e.g., a random noise vector with zero mean and extremely small variance) with its intensity strictly limited to a first intensity range on each dimension of the feature vector. By setting the first intensity range very small, it can be ensured that the perturbed new feature vector (i.e., the enhanced image features) remains closely distributed within the neighborhood of the original feature vector.

[0056] By introducing this controllable, minute randomness, we can simulate the natural fluctuations that may exist during feature extraction, thereby generating a set of highly similar but not completely identical candidate features for the subsequent model matching process. This improves the robustness of the matching process and avoids mismatches due to minor deviations in features.

[0057] 24. Apply a second type of enhancement transformation to the second type of image features whose confidence level is lower than a preset low threshold; the second type of enhancement transformation is a random perturbation within a second intensity range applied to the image features in the feature space, wherein the lower limit of the second intensity range is greater than the upper limit of the first intensity range.

[0058] This step involves extensive exploration within the feature space for low-confidence second-class image features.

[0059] Specifically, when the confidence level of the extracted image features is lower than a preset low threshold, it indicates that the location of the feature vector is unreliable and may be far from its ideal, true feature region. In this case, the applied second-type enhancement transformation is defined as: superimposing a random perturbation with an intensity falling within a second intensity range on each dimension of the feature vector. The lower limit of the second intensity range is greater than the upper limit of the first intensity range, meaning the perturbation intensity of the second-type transformation is significantly higher than that of the first-type transformation. This strong perturbation maps the original unreliable feature vector to a broader region in the feature space.

[0060] This large-scale, exploratory random sampling significantly increases the probability that one or more perturbed new feature vectors will accidentally fall into a feature region that a specific model excels at. This approach is a powerful remedial mechanism for low-quality features, compensating for insufficient accuracy with breadth.

[0061] 25. The enhanced image features are generated from the original image features through the random perturbation.

[0062] Through the aforementioned subtle fine-tuning and intense broad exploration—that is, by applying multiple random perturbations—several enhanced image features are derived from the original image features. These enhanced image features collectively constitute multiple possible interpretations of the original uncertain image in the feature space, facilitating high-probability, accurate matching in a dedicated model library.

[0063] This implementation achieves intelligent optimization of image features by applying a differentiated random perturbation strategy in the feature space based on confidence level. Specifically, small-range perturbations are used for fine-tuning of high-confidence features, enhancing the robustness of the matching process while maintaining feature accuracy; large-range perturbations are used for extensive exploration of low-confidence features, significantly increasing the probability of low-quality features matching the correct dedicated model. This adaptive enhancement mechanism at the feature level effectively overcomes the limitations of traditional image enhancement methods when dealing with uncertain features. It avoids excessive distortion of reliable features and effectively remedies unreliable features, providing richer and more reliable feature representations for subsequent model matching, thereby improving the accuracy and adaptability of yarn direction detection under complex working conditions.

[0064] like Figure 2 As shown, this embodiment of the invention discloses a yarn bobbin orientation detection system 200 based on image visual recognition technology. The system 200 includes: a model building unit 2001, which pre-builds a dedicated model library including multiple target detection models. Each target detection model is trained using a corresponding set of dedicated image data. The yarn end morphological features and yarn bobbin background features corresponding to each dedicated image data are similar.

[0065] The feature enhancement processing unit 2002 acquires a real-time image of the yarn package to be tested, extracts image features and their confidence levels from the real-time image, the image features including yarn end morphological features and yarn package background features; and performs enhancement transformation processing on the image features according to the confidence levels to obtain several enhanced image features.

[0066] The model matching and synthesis unit 2003 matches multiple relevant target detection models from the dedicated model library based on the enhanced image features, and derives a dynamic composite model based on each target detection model.

[0067] The recognition and reasoning unit 2004 uses the dynamic composite model to process the real-time image, obtains the predicted bounding box of the yarn end region, and determines the direction of the yarn bobbin based on the position of the predicted bounding box in the real-time image.

[0068] As an example, the model building unit 2001 is configured to: collect yarn package images that meet the quantity conditions, and classify all yarn package images according to the morphological features of the yarn ends and the background features of the yarn packages; wherein, the morphological features of the yarn ends include the yarn ends being pressed into the yarn layer, hanging naturally, or attached to the surface of the yarn packages in any shape; the background features of the yarn packages include different lighting conditions and the texture of the yarn package surface; grouping yarn package images with similar morphological features of the yarn ends and similar background features of the yarn packages into a group to form a set of dedicated image data, and using the set of dedicated image data to train an object detection model.

[0069] As an example, the feature enhancement processing unit 2002 is configured to: evaluate the saliency of the image features based on the degree of clustering in the feature space, wherein the degree of clustering is determined by calculating the distance between the image features and the cluster centers of similar features; and evaluate the reliability of the image features based on the stability of the image features in adversarial verification, wherein the stability is determined by calculating the degree of change of the image features after adding a small perturbation to the real-time image; and calculate the confidence level of the image features based on a weighted combination of the saliency and the reliability.

[0070] As an example, the feature enhancement processing unit 2002 is further configured to: determine the intensity and number of minute perturbations based on the image quality score of the real-time image and the neighborhood density of the image features in the feature space; add the determined minute perturbations to the real-time image to generate perturbed images; re-extract perturbed image features from each perturbed image; and evaluate the reliability of the image features by calculating the degree of difference between each perturbed image feature and the original image features.

[0071] As an example, the feature enhancement processing unit 2002 is further configured to: apply a first type of enhancement transformation to the first type of image features with a confidence level higher than a preset high threshold; the first type of enhancement transformation is a random perturbation within a first intensity range applied to the image features in the feature space; apply a second type of enhancement transformation to the second type of image features with a confidence level lower than a preset low threshold; the second type of enhancement transformation is a random perturbation within a second intensity range applied to the image features in the feature space, the lower limit of the second intensity range being greater than the upper limit of the first intensity range; and generate the plurality of enhanced image features from the original image features through the random perturbation.

[0072] Although the invention has been specifically shown and described with reference to preferred embodiments, those skilled in the art will understand that various modifications in form and detail may be made without departing from the spirit and scope of the invention. Accordingly, the disclosed invention should be considered merely illustrative and limited only by the scope specified in the appended claims.

Claims

1. A method for detecting the yarn bobbin orientation based on image visual recognition technology, characterized in that, The process includes the following steps: First, a dedicated model library containing multiple target detection models is pre-constructed, with each model trained using a corresponding set of dedicated image data. The morphological features of the yarn end and the background features of the yarn package corresponding to each set of dedicated image data are similar. Second, a real-time image of the yarn package to be tested is acquired, and image features and their confidence scores are extracted from the real-time image. These image features include yarn end morphological features and yarn package background features. Third, the image features are enhanced based on the confidence scores to obtain several enhanced image features. Fourth, multiple relevant target detection models are matched from the dedicated model library based on the enhanced image features, and a dynamic composite model is constructed based on each target detection model. Fifth, the dynamic composite model is used to process the real-time image to obtain a predicted bounding box for the yarn end region, and the direction of the yarn package is determined based on the position of the predicted bounding box in the real-time image.

2. The method for detecting yarn bobbin orientation based on image visual recognition technology according to claim 1, characterized in that: Each target detection model is trained using a corresponding set of dedicated image data, including: collecting yarn package images that meet the quantity requirements, and classifying all yarn package images based on yarn end morphological features and yarn package background features; wherein, the yarn end morphological features include yarn ends being pressed into the yarn layer, hanging naturally, or attached to the yarn package surface in any shape; the yarn package background features include different lighting conditions and yarn package surface texture; yarn package images with similar yarn end morphological features and similar yarn package background features are grouped together to form a set of dedicated image data, and a target detection model is trained using this set of dedicated image data.

3. The method for detecting yarn bobbin orientation based on image visual recognition technology according to claim 1, characterized in that: Extracting the confidence level of the image features from the real-time image includes: assessing the saliency of the image features based on their clustering degree in the feature space, wherein the clustering degree is determined by calculating the distance between the image features and the cluster centers of similar features; and assessing the reliability of the image features based on their stability in adversarial verification, wherein the stability is determined by calculating the degree of change of the image features after adding a small perturbation to the real-time image; and calculating the confidence level of the image features based on a weighted combination of the saliency and the reliability.

4. The method for detecting yarn bobbin orientation based on image visual recognition technology according to claim 3, characterized in that: The reliability of the image features is evaluated based on their stability in adversarial verification, including: determining the intensity and number of small perturbations based on the image quality score of the real-time image and the neighborhood density of the image features in the feature space; adding the determined small perturbations to the real-time image to generate perturbed images; re-extracting the perturbed image features from each perturbed image; and evaluating the reliability of the image features by calculating the degree of difference between each perturbed image feature and the original image features.

5. The method for detecting yarn bobbin orientation based on image visual recognition technology according to claim 4, characterized in that: The image features are enhanced based on the confidence level to obtain several enhanced image features, including: applying a first type of enhancement transformation to a first type of image features with a confidence level higher than a preset high threshold; the first type of enhancement transformation is a random perturbation within a first intensity range applied to the image features in the feature space; applying a second type of enhancement transformation to a second type of image features with a confidence level lower than a preset low threshold; the second type of enhancement transformation is a random perturbation within a second intensity range applied to the image features in the feature space, the lower limit of the second intensity range being greater than the upper limit of the first intensity range; and generating the several enhanced image features from the original image features through the random perturbation.

6. A yarn bobbin orientation detection system based on image visual recognition technology, characterized in that, The system includes: a model building unit, which pre-builds a dedicated model library containing multiple target detection models, each of which is trained using a corresponding set of dedicated image data; wherein the morphological features of the yarn end and the background features of the yarn package corresponding to each dedicated image data are similar; a feature enhancement processing unit, which acquires a real-time image of the yarn package to be tested, extracts image features and their confidence scores from the real-time image, the image features including yarn end morphological features and yarn package background features; and performs enhancement transformation processing on the image features according to the confidence scores to obtain several enhanced image features; a model matching and synthesis unit, which matches multiple related target detection models from the dedicated model library according to the enhanced image features, and constructs a dynamic composite model based on each target detection model; and a recognition and inference unit, which processes the real-time image using the dynamic composite model to obtain a predicted bounding box of the yarn end region, and determines the direction of the yarn package based on the position of the predicted bounding box in the real-time image.

7. The yarn bobbin orientation detection system based on image visual recognition technology according to claim 6, characterized in that: The model building unit is configured to: collect yarn package images that meet the quantity requirements, and classify all yarn package images according to the morphological features of the yarn ends and the background features of the yarn packages; wherein, the morphological features of the yarn ends include the yarn ends being pressed into the yarn layer, hanging naturally, or attached to the surface of the yarn packages in any shape; the background features of the yarn packages include different lighting conditions and the texture of the yarn package surface; grouping yarn package images with similar morphological features of the yarn ends and similar background features of the yarn packages into a group to form a set of dedicated image data, and using the set of dedicated image data to train an object detection model.

8. The yarn bobbin orientation detection system based on image visual recognition technology according to claim 6, characterized in that: The feature enhancement processing unit is configured to: evaluate the saliency of the image features based on the degree of clustering in the feature space, wherein the degree of clustering is determined by calculating the distance between the image features and the cluster centers of similar features; and evaluate the reliability of the image features based on the stability of the image features in adversarial verification, wherein the stability is determined by calculating the degree of change of the image features after adding a small perturbation to the real-time image; and calculate the confidence level of the image features based on a weighted combination of the saliency and the reliability.

9. A yarn bobbin orientation detection system based on image visual recognition technology according to claim 8, characterized in that: The feature enhancement processing unit is further configured to: determine the intensity and number of minute perturbations based on the image quality score of the real-time image and the neighborhood density of the image features in the feature space; add the determined minute perturbations to the real-time image to generate perturbed images; re-extract perturbed image features from each perturbed image; and evaluate the reliability of the image features by calculating the degree of difference between each perturbed image feature and the original image features.

10. A yarn bobbin orientation detection system based on image visual recognition technology according to claim 9, characterized in that: The feature enhancement processing unit is further configured to: apply a first type of enhancement transformation to the first type of image features with a confidence level higher than a preset high threshold; the first type of enhancement transformation is a random perturbation within a first intensity range applied to the image features in the feature space; apply a second type of enhancement transformation to the second type of image features with a confidence level lower than a preset low threshold; the second type of enhancement transformation is a random perturbation within a second intensity range applied to the image features in the feature space, the lower limit of the second intensity range being greater than the upper limit of the first intensity range; and generate the plurality of enhanced image features from the original image features through the random perturbation.

Images