An underwear gluing image recognition method and system based on high-precision positioning

Abstract

The present application relates to the technical field of underwear gluing image recognition, and particularly relates to a kind of underwear gluing image recognition method and system based on high-precision positioning, the method comprises the following steps: obtaining no trace underwear image data;The no trace underwear image data is divided into multiple local regions, and local brightness feature information is obtained;According to the local brightness feature information of each local region, the edge detection threshold of each local region is determined;Edge detection operator is applied to each multi-resolution image, and the edge graph of each multi-resolution image is extracted in combination with the edge detection threshold of each local region;Fusion each edge graph, and obtain preliminary edge contour;The preliminary edge contour is subjected to morphological processing, smoothing processing and connection processing, and refined gluing contour is obtained;The refined gluing contour is converted into coordinate point sequence, and the coordinate point sequence is output as gluing path instruction.The above is helpful to improve product quality and production efficiency.

Description

Technical Field

[0001] This invention relates to the technical field of image recognition of adhesive coating on underwear, and specifically to a method and system for image recognition of adhesive coating on underwear based on high-precision positioning. Background Technology

[0002] In automated production, especially in the manufacturing of seamless underwear where precision and appearance are paramount, accurate adhesive application is crucial for ensuring product functionality and comfort. This process typically relies on a sophisticated vision-guided system that uses an industrial camera to capture images of the underwear fabric, identify the adhesive application area, and translate the results into motion commands for a robotic arm to guide the nozzles for precise adhesive application. However, in the long-term, high-intensity industrial environment, LED lighting systems experience gradual and non-uniform light attenuation. This results in uneven light intensity distribution on the seamless underwear surface within the industrial camera's field of view, manifesting as gradually darkening areas and irregular brightness differences, thus reducing the quality of image recognition. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned shortcomings by proposing a method and system for recognizing images of adhesive coating on underwear based on high-precision positioning.

[0004] The present invention adopts the following technical solution: A method for recognizing adhesive coating images of underwear based on high-precision positioning, the method comprising the following steps: Acquire image data of seamless underwear; The seamless underwear image data is divided into multiple local regions, and the average brightness value, brightness standard deviation and brightness gradient of each local region are calculated to obtain local brightness feature information. Based on the local brightness feature information of each local region, determine the edge detection threshold of each local region; Multi-resolution processing is performed on seamless underwear image data to generate multiple multi-resolution images. An edge detection operator is applied to each multi-resolution image, and the edge map of each multi-resolution image is extracted by combining the edge detection threshold of each local region. By fusing the edge images, a preliminary edge outline is obtained; The initial edge contour is subjected to morphological processing, smoothing, and connection processing to obtain a refined adhesive coating contour. The refined adhesive application profile is converted into a sequence of coordinate points, and the sequence of coordinate points is output as an adhesive application path instruction.

[0005] This technical solution effectively addresses the impact of uneven and progressively decaying illumination on image recognition accuracy in industrial environments. Through multi-resolution processing, local brightness feature analysis, and adaptive edge detection, it achieves accurate recognition of the adhesive application contour of seamless underwear and converts it into high-precision adhesive application path instructions. This solves the problem of accumulated adhesive application errors caused by image recognition deviations in existing technologies, significantly improving product quality and production efficiency.

[0006] This application also discloses a high-precision positioning-based image recognition system for underwear adhesive application, applied to the aforementioned high-precision positioning-based image recognition method for underwear adhesive application. The system includes: The acquisition module acquires image data of seamless underwear; The calculation module divides the seamless underwear image data into multiple local regions and calculates the average brightness value, brightness standard deviation, and brightness gradient of each local region to obtain local brightness feature information. The determination module determines the edge detection threshold for each local region based on the local brightness feature information of each local region. The extraction module performs multi-resolution processing on the seamless underwear image data to generate multiple multi-resolution images. It then applies an edge detection operator to each multi-resolution image and, in conjunction with the edge detection thresholds for each local region, extracts the edge map of each multi-resolution image. The fusion module merges the edge images to obtain a preliminary edge contour. The processing module performs morphological processing, smoothing, and connection processing on the initial edge contour to obtain a refined adhesive coating contour. The output module converts the refined adhesive application contour into a sequence of coordinate points and outputs the sequence of coordinate points as an adhesive application path instruction.

[0007] This application provides a system for implementing the above-mentioned image recognition method through this technical solution. Through modular design, it can efficiently and stably perform various functions such as image data acquisition, feature calculation, edge extraction, fusion, processing and output, providing hardware and software support for the precise application of adhesive to seamless underwear, thereby realizing high-precision adhesive application in automated production.

[0008] This application provides high-precision adhesive application path instructions, significantly improving the adhesive application quality of seamless underwear, reducing production losses and resource waste, overcoming production defects in the prior art, and has significant practical value and technological progress.

[0009] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0010] Figure 1 This is a flowchart of a method for recognizing images of adhesive coating on underwear based on high-precision positioning according to the present invention. Figure 2 This is a schematic diagram of the structure of an image recognition system for underwear adhesive application based on high-precision positioning according to the present invention. Detailed Implementation

[0011] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated in advance. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0012] This embodiment provides a method and system for recognizing images of adhesive coating on underwear based on high-precision positioning, combined with... Figure 1 and Figure 2 As shown.

[0013] refer to Figure 1 A method for recognizing adhesive coating images of underwear based on high-precision positioning, the method comprising the following steps: Acquire image data of seamless underwear; The seamless underwear image data is divided into multiple local regions, and the average brightness value, brightness standard deviation and brightness gradient of each local region are calculated to obtain local brightness feature information. Based on the local brightness feature information of each local region, determine the edge detection threshold of each local region; Multi-resolution processing is performed on seamless underwear image data to generate multiple multi-resolution images. An edge detection operator is applied to each multi-resolution image, and the edge map of each multi-resolution image is extracted by combining the edge detection threshold of each local region. By fusing the edge images, a preliminary edge outline is obtained; The initial edge contour is subjected to morphological processing, smoothing, and connection processing to obtain a refined adhesive coating contour. The refined adhesive application profile is converted into a sequence of coordinate points, and the sequence of coordinate points is output as an adhesive application path instruction.

[0014] The term "seamless underwear image data" refers to the raw image information acquired by imaging equipment such as industrial cameras, used to identify adhesive-coated areas. This data typically includes pixel brightness and color information. "Local region" refers to dividing the entire seamless underwear image data into several smaller, independent image sub-blocks. Analyzing these sub-blocks helps to handle local illumination differences more precisely. "Average brightness value," "brightness standard deviation," and "brightness gradient" are important statistical features describing the distribution and changes in brightness in local regions. They collectively constitute "local brightness feature information," used to assess the uniformity of local illumination and the salience of edges. "Edge detection threshold" is a key parameter used to distinguish between edge and non-edge regions in an image; its dynamic determination is crucial for accurate edge extraction. "Multi-resolution processing" refers to generating images of different resolutions from the original image to capture edge information at different scales. "Edge detection operator" is an algorithm used in image processing to identify areas of drastic brightness change (i.e., edges), such as the Canny operator and the Sobel operator. "Edge map" is a binary image obtained after edge detection, containing only image edge information. The "preliminary edge contour" is the initial contour obtained by fusing edge images of different resolutions, and may contain some noise or discontinuities. "Morphological processing," "smoothing processing," and "connection processing" are image post-processing techniques used to optimize the preliminary edge contour, making it more continuous and smooth, and removing noise, ultimately resulting in the "refined adhesive coating contour." The "coordinate point sequence" represents the refined adhesive coating contour as a series of ordered two-dimensional coordinate points. These points can be directly input as "adhesive coating path instructions" to a high-precision positioning robotic arm, guiding it to perform precise adhesive coating.

[0015] The implementation methods described in this application will elaborate on each step of the above method.

[0016] Firstly, there are several methods for acquiring image data of seamless underwear. For example, a standard industrial camera can be used to periodically photograph the seamless underwear on the production line, and the acquired image data can be directly transmitted to the image processing unit. Alternatively, a scanning device integrated into the production line can be used to scan the seamless underwear line by line to generate complete image data. Furthermore, the seamless underwear can be placed on a specific shooting platform manually, photographed using a high-resolution digital camera, and then the image data imported into the system for processing.

[0017] Secondly, regarding the division of seamless underwear image data into multiple local regions and the calculation of the average brightness value, brightness standard deviation, and brightness gradient of each local region to obtain local brightness feature information, the following methods can be adopted. One approach is to uniformly divide the image data into fixed-size grid-like local regions, for example, dividing the image into 16x16 pixel or 32x32 pixel blocks. For each local region, the average brightness value can be obtained by calculating the sum of the gray values ​​of all its pixels divided by the number of pixels. The brightness standard deviation can be obtained by calculating the square root of the average of the sum of the squares of the differences between the gray values ​​of each pixel and the average brightness value. The brightness gradient can be obtained by applying the Sobel or Prewitt operator to the local regions, calculating the gradients in the horizontal and vertical directions, and then synthesizing them. Another approach is to use an adaptive segmentation method, dynamically adjusting the size and shape of the local regions according to the complexity of the image content and the brightness variations. For example, smaller local regions can be used in areas with large brightness variations, while larger local regions can be used in areas with uniform brightness.

[0018] Next, regarding determining the edge detection threshold for each local region based on its local brightness feature information, there are several possible approaches. One approach is to directly look up the corresponding edge detection threshold from a pre-defined lookup table based on the calculated average brightness value and brightness standard deviation. For example, when the average brightness value and brightness standard deviation of a local region are low, a lower threshold may be needed to avoid missing edges; conversely, when the average brightness value and brightness standard deviation are high, a higher threshold may be needed to suppress noise. Another approach is to use a simple linear or nonlinear mapping function to directly map the local brightness feature information (such as the combination of average brightness value and brightness standard deviation) to a threshold range. For example, a baseline threshold can be set, and then the baseline threshold can be adjusted by weighting based on the deviation between the average brightness value of the local region and the global average brightness.

[0019] Then, in processing seamless underwear image data for multi-resolution, generating multiple multi-resolution images, applying edge detection operators to each multi-resolution image, and combining edge detection thresholds for each local region to extract edge maps for each multi-resolution image, the following methods can be used. One implementation is to downsample the original image using methods such as Gaussian pyramids or Laplacian pyramids to generate a series of images at different resolutions. For example, images at the original resolution, half resolution, and quarter resolution can be generated. Then, for each resolution image, the Canny edge detection operator is applied to its corresponding local region, and edge extraction is performed using a pre-determined edge detection threshold for that local region, thus obtaining edge maps at each resolution. Another implementation is to use frequency domain methods such as wavelet transform to decompose the image into different frequency components, each frequency component corresponding to edge information at different scales, thereby forming multi-resolution images.

[0020] Subsequently, the following methods can be used to fuse the edge maps to obtain the preliminary edge contour. One implementation is to perform a simple logical "OR" operation on the edge maps of all resolutions, meaning that if an edge is detected in any edge map of any resolution, it is considered part of the preliminary edge contour. For example, high-resolution, medium-resolution, and low-resolution edge maps can be superimposed pixel-wise, and any pixel that is 1 (representing an edge) in any edge map will also be 1 in the preliminary edge contour. Another implementation is to use a weighted average for fusion, assigning different weights based on the reliability or importance of edge maps of different resolutions. For example, high-resolution edge maps can be given higher weights to preserve details, while low-resolution edge maps can be given lower weights to provide overall structural information.

[0021] Next, there are several ways to perform morphological processing, smoothing, and connection processing on the initial edge contour to obtain a refined adhesive coating contour. One approach is to first perform a morphological closing operation on the initial edge contour to fill small gaps and holes, followed by an opening operation to remove small noise points and burrs. Then, a Gaussian filter or median filter can be used to smooth the contour to reduce jagged edges. Finally, a distance- or connectivity-based algorithm is used to connect broken edge segments to form a continuous, refined adhesive coating contour. Another approach is to use curve fitting methods, such as B-spline curves or Bézier curves, to smooth and connect the initial edge contour. First, key points are extracted from the initial edge contour, and then these key points are used to fit a smooth curve, thus obtaining the refined adhesive coating contour.

[0022] Finally, the following methods can be used to convert the refined adhesive application contour into a sequence of coordinate points and output this sequence as adhesive application path instructions. One approach is to use a contour tracking algorithm, such as the Moore-Neighbor or Suzuki-Abe algorithm, to extract pixel coordinates from the contour point by point and arrange them in a clockwise or counterclockwise direction to form a sequence of coordinate points. These coordinate point sequences are then sent to the robotic arm controller via standard communication protocols (such as TCP / IP or RS232) as motion instructions. Another approach is to vectorize the refined adhesive application contour, converting it into a geometric shape composed of straight line segments and arc segments. Then, the parameters of these geometric shapes (such as start point, end point, radius, etc.) are converted into a sequence of G-code or M-code instructions that the robotic arm can recognize.

[0023] The high-precision positioning-based image recognition method for underwear adhesive coating proposed in this application effectively solves the problem of decreased image recognition accuracy caused by uneven lighting in traditional adhesive coating processes through a series of innovative steps. The method first acquires seamless underwear image data, which forms the basis for all subsequent processing. Next, the image data is divided into multiple local regions, and the average brightness value, brightness standard deviation, and brightness gradient of each local region are calculated to obtain local brightness feature information. This step is one of the core innovations of this application; it can precisely capture the lighting changes in local areas of the image, rather than simply performing global processing. By analyzing the local brightness feature information, the system can dynamically determine the edge detection threshold of each local region according to the actual lighting conditions, avoiding the failure of fixed thresholds under uneven lighting.

[0024] Subsequently, this application performs multi-resolution processing on the seamless underwear image data to generate multiple multi-resolution images. This processing method enables the system to capture edge information at different scales; high-resolution images retain details, while low-resolution images provide overall structure, thereby improving the robustness of edge detection. Edge detection operators are applied to each multi-resolution image, and combined with edge detection thresholds for each local region, edge maps of each multi-resolution image are extracted, ensuring accurate edge extraction at different resolutions.

[0025] The key to this application lies in obtaining a preliminary edge contour by fusing various edge images. This fusion mechanism can comprehensively utilize the advantages of different resolutions, compensating for the shortcomings of single-resolution edge detection. For example, high resolution may be sensitive to noise, while low resolution may lose details. Through intelligent fusion, a more comprehensive and accurate preliminary contour can be obtained. Finally, morphological processing, smoothing, and connection processing are performed on the preliminary edge contour to obtain a refined adhesive application contour. These post-processing steps further optimize the quality of the contour, making it more continuous and smooth, and removing noise and unnecessary breaks, thereby generating a precise path that can be directly used for adhesive application. Finally, the refined adhesive application contour is converted into a sequence of coordinate points and output as an adhesive application path instruction, realizing the precise guidance of a high-precision positioning robotic arm.

[0026] The steps for fusing the various edge maps to obtain the preliminary edge contours include: Local brightness consistency analysis is performed on each edge map. The average brightness difference between the high-resolution edge map and the low-resolution edge map in each local region is calculated. Based on the average brightness difference, it is determined whether there is distortion caused by inconsistent illumination in each local region, and the results of the local brightness consistency analysis are obtained. Calculate the local contrast intensity and connectivity of edge pixels in each edge map, and mark weak edge regions in the high-resolution edge map based on the local contrast intensity and connectivity. Based on the labeling results, strong edge regions in high-resolution edge maps are given priority. For weak edge regions in high-resolution edge maps, medium-resolution edge information is used to supplement them and make up for the lack of detail in the high-resolution edge maps. For weak edges or missing regions in the medium-resolution edge map, the overall contour information in the low-resolution edge map is weighted and corrected based on the results of local brightness consistency analysis. The corrected overall contour information in the low-resolution edge map is then filled into the medium-resolution edge map to obtain the preliminary edge contour.

[0027] The local brightness consistency analysis of each edge map involves comparing the average brightness values ​​of edge maps at different resolutions in the same local area to assess whether there are brightness differences in that area due to uneven illumination or changes in imaging conditions. These differences can lead to inconsistent edge information representation at different resolutions, thus affecting the accuracy of fusion. By calculating the average brightness difference between high-resolution and low-resolution edge maps in each local area, this inconsistency can be quantified, and the presence of distortion caused by inconsistent illumination in each local area can be determined, thus yielding the results of the local brightness consistency analysis.

[0028] Furthermore, local contrast intensity and connectivity are calculated for edge pixels in each edge map to assess the reliability and integrity of the edges. Local contrast intensity reflects the distinction between the edge and the surrounding background; high contrast generally indicates a more reliable edge. Connectivity describes the degree of connection between edge pixels; continuous edge segments are usually more meaningful than isolated pixels. By comprehensively considering local contrast intensity and connectivity, weak edge regions in high-resolution edge maps can be effectively labeled. These regions may have unclear or incomplete edge information due to noise, blurring, or insufficient lighting.

[0029] Based on the above labeling results, the scheme in this application prioritizes strong edge regions in the high-resolution edge image. This is because high-resolution images have a natural advantage in detail representation, and their strong edge regions typically correspond to key parts of the adhesive outline, exhibiting high accuracy and reliability. For weak edge regions in the high-resolution edge image, medium-resolution edge information is used to supplement them. Medium-resolution images achieve a balance between detail and overall structure, and their edge information can effectively compensate for potential detail loss or instability in weak edge regions of the high-resolution edge image.

[0030] Furthermore, for weak edges or missing regions in the medium-resolution edge map, the proposed solution combines local brightness consistency analysis results to perform weighted correction on the overall contour information of the low-resolution edge map. Although low-resolution images have fewer details, their overall contour information is usually more stable and less affected by noise and local illumination variations. By performing weighted correction on the low-resolution edge map using local brightness consistency analysis results, distortion caused by illumination inconsistencies can be effectively compensated during the fusion process. Finally, the corrected overall contour information from the low-resolution edge map is filled into the medium-resolution edge map, resulting in a preliminary edge contour that contains rich details and has good overall consistency.

[0031] This application's solution, by introducing local brightness consistency analysis of multi-resolution edge maps, can effectively identify and address potential uneven illumination issues in images, avoiding edge information distortion caused by illumination differences. Simultaneously, by performing local contrast intensity and connectivity analysis on edge pixels, it can intelligently distinguish between strong and weak edges, thus prioritizing the preservation of reliable details from high-resolution images during the fusion process, while utilizing the advantages of medium- and low-resolution images for supplementation and correction. This hierarchical and adaptive fusion strategy allows edge information from different resolutions to complement each other, effectively solving problems such as detail loss, noise introduction, or incomplete contours that may result from traditional simple fusion methods.

[0032] In some preferred embodiments, it is assumed that adhesive application path recognition is being performed on a seamless bra. During the photographing of the bra, uneven ambient lighting resulted in shadows in some areas of the image, causing some adhesive edges to appear as weak or blurred edges. Using only a single resolution or a simple edge fusion method may lead to inaccurate or broken adhesive contour recognition in the shadowed areas. The solution in this application first performs local brightness consistency analysis on high, medium, and low resolution edge maps to identify brightness differences in shadowed areas and determine if there is distortion caused by inconsistent lighting in these areas. Next, the local contrast intensity and connectivity of edge pixels in the high-resolution edge map are calculated to accurately mark weak edges within the shadowed areas. During fusion, for areas with sufficient lighting and clear edges, strong edge information from the high-resolution edge map is preferentially adopted to preserve fine details. For weak edges within the shadowed areas, medium-resolution edge information is used as a supplement, because medium-resolution images can smooth noise to some extent and provide more continuous edges. Finally, for weak edges or missing areas that may still exist in the shadow regions of the medium-resolution edge map, the overall contour information in the low-resolution edge map is weighted and corrected based on the previously obtained local brightness consistency analysis results. For example, higher weights are given to the low-resolution edge map in areas with lower brightness to ensure the integrity and accuracy of the overall contour. The corrected overall contour information from the low-resolution edge map is then incorporated into the medium-resolution edge map. Through this multi-level, adaptive fusion process, even under complex conditions of uneven lighting, a high-precision and highly complete preliminary adhesive coating contour can be obtained, laying a solid foundation for subsequent fine processing.

[0033] This application further proposes steps for calculating the local contrast intensity and connectivity of edge pixels in each edge map, including: Texture feature analysis is performed on local regions of each edge map to obtain texture feature information; Based on texture feature information, the local contrast intensity threshold of local regions in each edge map is corrected; Calculate the local contrast intensity for edge pixels in each edge map, and determine the edge pixels based on the corrected local contrast intensity threshold; Connectivity analysis is performed on edge pixels to obtain connected edge segments; The length, rate of curvature change, and width consistency of connected edge segments are analyzed to obtain the geometric feature analysis results. Based on the geometric feature analysis results, connected edge segments that do not conform to the glue coating edge characteristics are removed, and the remaining connected edge segments are obtained. The stability of the remaining connected edge segments is evaluated, and the stability evaluation results are obtained. Based on the stability assessment results, unstable edge segments are removed.

[0034] Specifically, texture feature analysis refers to quantifying the texture attributes of local image regions, such as roughness, contrast, and directionality, by calculating the gray-level co-occurrence matrix features, local binary patterns, or Gabor filter responses of each local region, thereby obtaining texture feature information. This texture feature information characterizes the visual properties of local regions and helps distinguish between smooth, textured, and edge regions. Furthermore, based on the obtained texture feature information, the local contrast intensity threshold of each edge map region can be adaptively adjusted. For example, in regions with complex textures, the threshold can be appropriately increased to avoid misidentifying textures as edges; while in smooth regions, the threshold can be appropriately decreased to capture subtle coated edges. After determining the edge pixels, connectivity analysis connects adjacent edge pixels to form connected edge segments. Connectivity analysis can employ eight-neighbor or four-neighbor connection rules to group pixels with the same edge attributes into the same segment.

[0035] Subsequently, geometric feature analysis is performed on these connected edge segments. Specifically, length analysis is used to exclude uncoated edge segments that are too short or too long; curvature change rate analysis is used to identify the smoothness of the edges, as coated edges typically have relatively smooth curves; and width consistency analysis is used to ensure that the width of the edge segments remains stable within a certain range to conform to the physical characteristics of the coated lines. Through these geometric feature analyses, geometric feature analysis results can be obtained, and connected edge segments that do not conform to the preset coated edge characteristics can be eliminated, thus obtaining the remaining connected edge segments. Finally, the stability of the remaining connected edge segments is evaluated. The stability evaluation can be based on factors such as the gradient intensity of the edge, the consistency of local contrast, and the persistence of the edge at different scales. For example, the robustness can be evaluated by analyzing the changes of the edge under small perturbations. Based on the stability evaluation results, unstable edge segments that are prone to breakage or disappearance under image noise or small changes can be further eliminated, thus ensuring that the final edge segments are reliable and stable.

[0036] This application's solution significantly improves the accuracy and robustness of calculating local contrast intensity and connectivity for edge pixels by introducing texture feature analysis, adaptive threshold correction, multi-dimensional geometric feature analysis, and stability evaluation. First, texture feature analysis provides a deeper understanding of the intrinsic characteristics of local image regions, making the correction of local contrast intensity thresholds more targeted and avoiding misjudgments that might occur with fixed thresholds in complex backgrounds. Second, connectivity analysis organizes discrete edge pixels into meaningful connected edge segments, laying the foundation for subsequent refined selection. Third, by analyzing the length, rate of curvature change, and width consistency of connected edge segments, it is possible to effectively distinguish real coated edges from non-coated textures or noise from a geometric perspective, for example, eliminating short, curved, or irregular edge segments that do not conform to the typical geometric characteristics of coated lines. Finally, stability evaluation further ensures the reliability of the identified edge segments, eliminating unstable edges that are susceptible to interference, thus making the final edge information more accurate and reliable. It is precisely because of these multi-level refined processing that weak edge regions in high-resolution edge maps can be more accurately marked, providing high-quality input for subsequent multi-resolution edge map fusion.

[0037] In some preferred embodiments, it is assumed that there are some fine textures formed by the fabric's own weave in the seamless underwear image data. These textures may exhibit similar contrast intensities to glued edges in local areas. If only a fixed local contrast intensity threshold is used to determine edge pixels, these fabric textures may be incorrectly identified as edge pixels and thus marked as weak edge regions, leading to noise during subsequent fusion. The solution of this application addresses this problem in the following way: First, texture feature analysis is performed on these local regions, such as calculating their local binary pattern features, to find that they have periodic or random texture patterns rather than continuous edge structures. Based on this texture feature information, the system adaptively adjusts the local contrast intensity threshold of the local region, appropriately increasing it in the textured region, thereby avoiding misidentification of fabric textures as edges. Second, even if some texture pixels are initially identified as edge pixels due to their high contrast, after subsequent connectivity analysis, the connected edge segments formed by these texture pixels usually exhibit geometric features that do not conform to glued edges, such as being too short in length, having an excessively large rate of curvature change, or having inconsistent width. For example, the edges of adhesive coating are typically relatively smooth curves of a certain length, while the "edges" formed by fabric textures are often scattered, irregular short line segments. By performing length, rate of curvature change, and width consistency analysis on these connected edge segments, these non-compliant connected edge segments can be eliminated based on preset geometric feature parameters of the adhesive coating edges. Finally, the stability of the remaining connected edge segments is evaluated. Those "edges" formed accidentally by fabric textures often exhibit instability under minor image changes or noise interference, easily breaking or disappearing. Through stability evaluation, these unstable edge segments will be further eliminated. Through the above series of refined processing, this application can effectively and accurately separate the real adhesive coating edges from complex fabric textures and noise, ensuring more accurate labeling of weak edge areas, thus providing high-quality input for subsequent edge map fusion, and ultimately improving the accuracy of adhesive coating path instructions.

[0038] The steps for performing local brightness consistency analysis on each edge map further include: Based on the average brightness difference and the known optical reflection characteristic curves of seamless underwear fabric, the average brightness difference is mapped to the edge information distortion value. Based on the degree of edge information distortion, the weight distribution of high-resolution edge maps and low-resolution edge maps during fusion is dynamically adjusted. Based on the distortion value of edge information, the local contrast intensity threshold of local regions in each edge map is corrected according to the local brightness consistency analysis.

[0039] Specifically, the average brightness difference refers to the average brightness difference between high-resolution and low-resolution edge images in various local areas, reflecting the illumination consistency of images of different resolutions in the same area. The known optical reflectance characteristic curve of seamless underwear fabric refers to a curve obtained in advance through experiments or measurements, describing the characteristics of light reflection by a specific seamless underwear fabric under different lighting conditions. This curve can characterize the fabric's absorption, reflection, and scattering behavior of light, such as reflectivity at different incident angles or wavelengths. Mapping the average brightness difference to an edge information distortion value can be achieved by establishing a mapping function or lookup table, taking the calculated average brightness difference as input, combining it with the fabric's optical reflectance characteristic curve, and outputting a quantified edge information distortion value. The larger this value, the greater the impact of inconsistent illumination on the edge information of that local area, and the more severe the distortion.

[0040] The dynamic adjustment of the weight allocation between high-resolution and low-resolution edge maps during fusion refers to assigning a contribution ratio to each when fusing them. When the edge information distortion level is high, the weight of edge maps more affected by distortion (e.g., high-resolution edge maps) can be appropriately reduced, while the weight of edge maps less affected by distortion (e.g., low-resolution edge maps) can be increased, and vice versa. The local contrast intensity threshold of each edge map's local region is corrected based on local brightness consistency analysis. This local contrast intensity threshold is used to determine whether a pixel is an edge point. When the edge information distortion level is high, this threshold may need to be adjusted; for example, the threshold may be appropriately widened or tightened in distorted regions to avoid false or missed edge detections.

[0041] This application's solution quantifies the average brightness difference between high-resolution and low-resolution edge images as an edge information distortion value by introducing the known optical reflectance characteristic curve of seamless underwear fabric. It is precisely this precise quantification of distortion that allows the system to more meticulously understand the impact of lighting inconsistencies on edge images of different resolutions. Based on this, the distortion value is used to dynamically adjust the weight allocation between high-resolution and low-resolution edge images during fusion. For example, when the edge information distortion value is high in a certain local area, it indicates that the edge information in that area may be severely affected by uneven lighting. In this case, the weight of the high-resolution edge image in that area can be reduced, and more reliance can be placed on the relatively stable low-resolution edge image, thus avoiding the introduction of distortion information into the initial edge contour. Simultaneously, this distortion value is also used to correct the local contrast intensity threshold of each edge image based on local brightness consistency analysis. By dynamically adjusting the threshold according to the distortion degree, the edge detection process can better adapt to local lighting changes, avoiding excessive false edges or missing true edges in areas of inconsistent lighting, thereby improving the accuracy and robustness of edge detection.

[0042] In some preferred embodiments, assuming that during the acquisition of seamless underwear image data, uneven ambient lighting causes significant brightness differences in a local area of ​​the image, for example, the brightness of the high-resolution edge image in this area is significantly higher than that of the low-resolution edge image. First, the system calculates the average brightness difference in this local area. Then, combining this with the pre-measured optical reflectance characteristic curve of the seamless underwear fabric (e.g., a microfiber fabric of a specific material), the system maps this average brightness difference to a high edge information distortion value, such as 0.8 (range 0-1). Based on this high distortion value, the system dynamically adjusts the fusion weights of the high-resolution and low-resolution edge images in this local area. Specifically, the weight of the high-resolution edge image may be reduced from the default 0.6 to 0.4, while the weight of the low-resolution edge image may be increased from the default 0.4 to 0.6, to reduce pseudo-edges that may exist in the high-resolution edge image due to uneven lighting or to enhance the stability of the low-resolution edge image. Simultaneously, this higher distortion value is also used to correct the local contrast intensity threshold of this local area. For example, if the original threshold is T, the system may adjust it to T'=T based on the distortion level. (1-0.8) (k), where k is an adjustment coefficient, thereby employing a more conservative threshold in distorted areas to avoid misjudging brightness differences as edges. This dynamic adjustment ensures that the extracted initial edge contours are more accurate and reliable, even under inconsistent lighting conditions.

[0043] The steps to obtain local brightness feature information include: Noise type identification is performed in each local area; Based on the identified noise type, select and apply the corresponding adaptive noise suppression filter to preprocess each local area; Statistical analysis is performed on the pixel grayscale values ​​of each preprocessed local region to calculate the average brightness value and brightness standard deviation of each local region. Apply edge detection operators to each preprocessed local region and calculate the brightness gradient of each local region; Local brightness characteristic information is obtained based on the average brightness value, brightness standard deviation, and brightness gradient of each local area.

[0044] Noise type identification refers to determining the type of noise primarily affecting a local area of ​​an image by analyzing its statistical characteristics, texture features, or frequency domain characteristics. Common noise types include Gaussian noise, salt-and-pepper noise, and Poisson noise. Accurately identifying the noise type is a prerequisite for selecting an appropriate denoising method.

[0045] An adaptive noise suppression filter is a filter whose filtering parameters or filtering strategy can be adjusted according to local image characteristics (such as noise intensity, texture complexity, etc.). For example, for Gaussian noise, an adaptive Gaussian filter or Wiener filter can be used; for salt-and-pepper noise, an adaptive median filter can be used. By selecting a filter that matches the noise type, noise can be removed more effectively while preserving image details to the maximum extent.

[0046] The average brightness value reflects the overall brightness level of a local area, while the standard deviation of brightness characterizes the dispersion or contrast of pixel brightness within that local area. After preprocessing, these statistics can more accurately reflect the true brightness characteristics of the image and reduce bias caused by noise.

[0047] The brightness gradient is a measure of the rate of change of pixel brightness, commonly used to indicate edge information in an image. Common edge detection operators include the Sobel operator, Prewitt operator, and Canny operator. Calculating the brightness gradient on a noise-suppressed image yields a clearer and more accurate edge response, avoiding false edges caused by noise.

[0048] Local brightness features are obtained by analyzing the average brightness value, standard deviation of brightness, and brightness gradient of each local region. These features together constitute a comprehensive description of the brightness characteristics of the local region, providing a reliable basis for subsequently determining the edge detection threshold.

[0049] This application's solution effectively addresses the impact of noise on the accuracy of brightness feature calculation in the original scheme by introducing noise type identification and adaptive noise suppression preprocessing before calculating local brightness feature information. Specifically, firstly, noise type identification is performed on each local region, enabling the system to specifically understand the noise characteristics present in the image. Since different types of noise have different statistical distributions and visual representations, different denoising strategies are required. Secondly, based on the identified noise type, a corresponding adaptive noise suppression filter is selected and applied to preprocess each local region. This adaptability ensures that the filter can be optimized and adjusted according to the local image content and noise characteristics, thereby effectively removing noise while preserving image details and edge information to the maximum extent, avoiding problems of over-smoothing or insufficient denoising. Based on this, the average brightness value, brightness standard deviation, and brightness gradient are calculated on the preprocessed image data. These calculation results are more accurate and stable, truly reflecting the brightness characteristics and edge intensity of local regions. Because this feature information is more reliable, the subsequent edge detection threshold determination is more precise, thus laying a solid foundation for the final refined extraction of the adhesive coating contour.

[0050] In some preferred embodiments, it is assumed that in the acquired seamless underwear image data, some local areas have obvious Gaussian noise due to insufficient lighting or sensor defects, while other local areas have salt-and-pepper noise due to data transmission errors or pixel damage.

[0051] First, the system divides the seamless underwear image data into multiple local regions. For each local region, noise type identification is performed. For example, by analyzing the pixel grayscale histogram of the local region, if the histogram shows a bell-shaped distribution and a large variance, it may be identified as Gaussian noise; if the histogram has peaks at 0 and 255, it may be identified as salt-and-pepper noise.

[0052] Once the noise type is identified, the system selects and applies the corresponding adaptive noise suppression filter. For example, for regions identified as Gaussian noise, an adaptive Gaussian filter can be applied, whose smoothing kernel size and intensity are adjusted according to the local noise level and texture complexity to avoid excessive blurring. For regions identified as salt-and-pepper noise, an adaptive median filter is applied, which effectively removes isolated anomalous pixels while preserving edge details.

[0053] After preprocessing, the image data of each local region becomes clearer, and noise is effectively suppressed. Based on this, the system calculates the average brightness value and standard deviation of brightness of these preprocessed local regions, and applies the Sobel operator to calculate the brightness gradient. For example, the average brightness value and standard deviation of a region that was originally heavily contaminated by Gaussian noise will be closer to its true value after denoising, and the brightness gradient will more accurately reflect the actual edge intensity, rather than a pseudo-gradient caused by noise.

[0054] Ultimately, based on the average brightness value, brightness standard deviation, and brightness gradient obtained after noise suppression, the system can obtain more accurate and robust local brightness feature information, providing a solid foundation for subsequent edge detection and adhesive contour extraction.

[0055] The steps for identifying the noise type in each local area include: Texture features are extracted from each local region to obtain texture feature parameters; Based on the texture feature parameters, the noise sensitivity of each local region is evaluated, and the noise sensitivity evaluation results are obtained. Based on the noise sensitivity assessment results, the pixel grayscale histogram of each local region is adaptively segmented to obtain a segmented histogram. Perform statistical characteristic analysis on each segment histogram, and calculate the skewness, kurtosis and entropy of each segment histogram; Based on the skewness, kurtosis, and entropy of each segment histogram, and combined with the noise sensitivity assessment results, the noise type of each local area is identified as Gaussian noise, salt-and-pepper noise, or Poisson noise.

[0056] The process involves extracting texture features from various local regions to quantify the visual characteristics of these regions, such as roughness, directionality, and contrast. Texture feature parameters can include gray-level co-occurrence matrix features, local binary mode features, or Gabor filter responses, which effectively describe the structural information of local image regions.

[0057] Furthermore, based on the extracted texture feature parameters, the noise sensitivity of each local region can be evaluated. The noise sensitivity evaluation results reflect the resistance or susceptibility of different texture regions to noise. For example, in regions with flat textures, small noises may cause significant visual changes, while in regions with complex textures, noise may be masked by the texture.

[0058] Based on this, and according to the noise sensitivity assessment results, the pixel grayscale histograms of each local region can be adaptively segmented. Adaptive segmentation refers to dynamically adjusting the histogram division method according to the characteristics of the local region, in order to better capture the pixel distribution characteristics within different grayscale ranges. This results in a segmented histogram.

[0059] Subsequently, statistical feature analysis was performed on each segment of the histogram, calculating the skewness, kurtosis, and entropy value of each segment. Skewness measures the symmetry of the histogram, kurtosis measures the sharpness of the histogram, and entropy measures the randomness or information content of the histogram. These statistical features can reflect the characteristics of pixel grayscale distribution from different dimensions, thus providing a quantitative basis for noise type identification.

[0060] Finally, based on the skewness, kurtosis, and entropy of each segmented histogram, and combined with the previously obtained noise sensitivity assessment results, the noise type of each local region can be accurately identified as Gaussian noise, salt-and-pepper noise, or Poisson noise. For example, Gaussian noise typically results in a bell-shaped histogram distribution, while salt-and-pepper noise will have spikes at both ends of the histogram, and the distribution of Poisson noise is related to the signal strength.

[0061] This application's solution provides a solid foundation for subsequent noise type identification by performing refined texture feature extraction and noise sensitivity assessment on local image regions. Specifically, obtaining texture feature parameters enables the system to understand the intrinsic structure of local image regions, while noise sensitivity assessment further considers the differences in the response of these structures to noise. By adaptively segmenting the pixel grayscale histogram, the pixel distribution characteristics within different grayscale ranges can be captured more accurately, avoiding the information loss that may result from a single global histogram. Subsequently, statistical features such as skewness, kurtosis, and entropy are calculated on the segmented histogram. These quantitative indicators can effectively reveal the unique patterns of different noise types in pixel distribution. For example, Gaussian noise tends to cause pixel values ​​to be symmetrically distributed around a certain mean, resulting in a histogram skewness close to zero and moderate kurtosis; salt-and-pepper noise manifests as discrete points with extremely high and low grayscale values, causing obvious peaks at both ends of the histogram; Poisson noise, due to its counting properties, has a variance correlated with the mean, causing the histogram shape to change with brightness. Ultimately, by combining these statistical characteristics and noise sensitivity assessment results, the system can distinguish and identify the specific noise type experienced by a local area with higher accuracy, thus providing a precise basis for the subsequent selection of adaptive noise suppression filters.

[0062] This application further proposes steps for assessing the noise sensitivity of each local area, including: Anomaly detection is performed on each local area to identify abnormal areas such as microbubbles, fiber agglomerates, or adhesive residue; When evaluating the noise sensitivity of each local region, pixel data from abnormal regions are excluded; Based on the texture feature parameters, the pixel grayscale distribution characteristics of each local region excluding abnormal areas are analyzed to obtain the masking effect parameters of the inherent texture of the fabric on noise. Based on the masking effect parameters, the noise sensitivity assessment weights for each local region are adjusted to obtain the noise sensitivity assessment results.

[0063] Specifically, anomaly detection in local areas refers to identifying regions in an image that significantly deviate from the normal fabric texture and brightness distribution using specific image processing algorithms. These anomaly regions may manifest as tiny circular bright spots (microbubbles), irregular clumps (fiber agglomerations), or blurred patches (adhesive residue). They are not image noise but rather defects in the product itself or residues from the manufacturing process. The purpose of identifying these anomaly regions is to distinguish them from genuine noise and prevent them from interfering with noise sensitivity assessment. When assessing the noise sensitivity of each local area, excluding pixel data from anomaly regions means that once an anomaly region is identified, its internal pixel data will not be included in subsequent noise sensitivity calculations. This ensures that the noise sensitivity assessment is based solely on normal fabric areas, thereby improving the accuracy of the assessment.

[0064] Furthermore, based on texture feature parameters, the pixel grayscale distribution characteristics of each local region after excluding anomalous areas are analyzed to obtain the masking effect parameter of the fabric's inherent texture on noise. This masking effect parameter refers to the visual perception or physical suppression ability of different fabric materials and texture structures against different types of noise. For example, some dense fabrics or those with complex patterns may have a strong masking effect on Gaussian noise, making the noise less visually noticeable; while for discrete noises such as salt-and-pepper noise, the masking effect may be weaker. By analyzing the pixel grayscale distribution characteristics after excluding anomalous areas, this masking effect can be quantified, for example, by calculating parameters such as texture contrast, roughness, and directionality, and combining them with empirical models or machine learning methods to infer the degree of masking for different noise types. Finally, based on the masking effect parameter, the noise sensitivity assessment weights of each local region are adjusted to obtain the noise sensitivity assessment results. This means that when comprehensively assessing noise sensitivity, the weight of a particular noise type in the assessment is dynamically adjusted according to the fabric's masking effect on that specific noise. If the fabric has a strong masking effect on a certain type of noise, the sensitivity weight to that noise may be appropriately reduced, and vice versa, so that the final noise sensitivity assessment result is more in line with the actual situation.

[0065] This application's solution introduces anomaly detection before noise sensitivity assessment, effectively identifying and eliminating non-noise-related interference factors in the image, such as microbubbles, fiber agglomerates, or adhesive residue. After these anomalous pixel data areas are excluded, the subsequent noise sensitivity assessment will no longer be misled by these local defects, thus making the assessment results more realistically reflect the actual noise level of each local area. Furthermore, by analyzing the pixel grayscale distribution characteristics after excluding anomalous areas, the masking effect of the fabric's inherent texture on noise can be quantified. The introduction of this masking effect parameter allows the noise sensitivity assessment to fully consider the influence of different fabric textures on noise representation; for example, certain dense or complex textured fabrics may have a natural suppression or masking effect on specific types of noise. Finally, the noise sensitivity assessment weights are dynamically adjusted according to this masking effect parameter, ensuring that the assessment results not only eliminate anomalous interference but also more accurately match the actual noise characteristics of different fabric regions, providing a more reliable foundation for subsequent noise type identification and adaptive filtering.

[0066] In some preferred embodiments, it is assumed that after segmenting the seamless underwear image data into local regions, microbubbles are detected in a certain local area. When assessing the noise sensitivity of this local area, the microbubble region is first identified using an anomaly detection algorithm, and its pixel data is excluded from the evaluation sample. Subsequently, texture feature analysis is performed on the remaining pixel data after excluding the bubble region, such as calculating the gray-level co-occurrence matrix or wavelet transform coefficients, to obtain the inherent texture information of the fabric. Based on this texture information, it can be analyzed that the fabric texture has a strong masking effect on Gaussian noise but a weak masking effect on salt-and-pepper noise, thus obtaining the corresponding masking effect parameter. According to this parameter, when calculating the noise sensitivity assessment result, the weight of the sensitivity to Gaussian noise can be appropriately reduced, while the weight of the sensitivity to salt-and-pepper noise can be increased. For example, if the original assessment result shows similar sensitivity to Gaussian noise and salt-and-pepper noise, after adjusting the masking effect parameter, it may be concluded that the region is more sensitive to salt-and-pepper noise, thus making it more likely to be identified as salt-and-pepper noise in subsequent noise type identification, and selecting a more suitable adaptive filter for preprocessing salt-and-pepper noise.

[0067] This application further proposes a step for detecting abnormal points in various local areas and identifying abnormal areas such as microbubbles, fiber agglomerates, or adhesive residues, including: Image data is obtained by acquiring images of each local area; Multispectral analysis was performed on the original image data to obtain image data in different spectral bands; Multispectral texture features are obtained by extracting texture features from image data in different spectral bands. Color features are extracted from image data in different spectral bands to obtain multispectral color features; Shape features are extracted from image data in different spectral bands to obtain multispectral shape features; Based on multispectral texture features, multispectral color features, and multispectral shape features, identify the anomaly types in local regions; Based on the anomaly type in the local area, select and apply the corresponding anomaly detection threshold; Based on the anomaly detection threshold, abnormal areas such as microbubbles, fiber agglomerates, or adhesive residues are identified.

[0068] Specifically, acquiring images of various local areas to obtain raw image data refers to obtaining the initial visual information of the seamless underwear to be tested through imaging equipment. Raw image data can be understood as a set of image pixels directly acquired from the sensor without any processing. Further, multispectral analysis is performed on the raw image data to obtain image data in different spectral bands. The purpose is to obtain richer information about the material's optical properties beyond the range of human vision. Multispectral analysis can reveal the reflection, absorption, or transmission characteristics of materials at different wavelengths, which are important for distinguishing different substances (such as fabrics, bubbles, fibers, and adhesives). Extracting texture features from the image data in different spectral bands to obtain multispectral texture features involves calculating the texture attributes of the image on multiple spectral channels, such as the gray-level co-occurrence matrix, local binary mode, or Gabor filter response, to capture the fine structural changes of abnormal areas under different spectra. In practical applications, color features are extracted from image data across different spectral bands to obtain multispectral color features. For example, the average brightness, chromaticity, and saturation of each spectral channel can be calculated, or descriptors such as color histograms and color moments can be used to quantify the color representation of abnormal regions across different spectra. Furthermore, shape features are extracted from image data across different spectral bands to obtain multispectral shape features, the purpose of which is to describe the geometric morphology of abnormal regions. For example, the area, perimeter, aspect ratio, roundness, and eccentricity of abnormal regions can be extracted. These features help distinguish regularly shaped bubbles from irregular fiber agglomerates or adhesive residues. Therefore, identifying the anomaly type of a local region based on multispectral texture features, multispectral color features, and multispectral shape features involves comprehensively utilizing the aforementioned multidimensional feature vectors and classifying each detected abnormal region using a classifier (such as a support vector machine, neural network, or decision tree) to determine whether it belongs to microbubbles, fiber agglomerates, or adhesive residues. Further, based on the anomaly type of the local region, a corresponding anomaly detection threshold is selected and applied, aiming to employ the most appropriate discrimination criteria for different types of anomalies. For example, for microbubbles, thresholds based on roundness and brightness might be used; for fiber agglomerations, thresholds based on texture roughness and length might be used; and for adhesive residue, thresholds based on color deviation and shape irregularity might be used. Ultimately, identifying anomalous regions such as microbubbles, fiber agglomerations, or adhesive residue based on the anomaly detection threshold means marking sets of pixels or connected regions that meet specific anomaly type threshold conditions as the corresponding anomalous regions and excluding them from subsequent noise sensitivity assessments.

[0069] This application's solution overcomes the limitations of traditional single-spectral imaging in anomaly detection by introducing multispectral analysis. Traditional methods may struggle to accurately distinguish different types of anomalies or misclassify normal textures as anomalies due to insufficient spectral information. It is precisely because different types of anomalies (such as microbubbles, fiber agglomerates, and adhesive residues) and normal fabrics possess unique optical reflection, absorption, or transmission characteristics in different spectral bands that multispectral imaging can acquire richer and more discriminative information. By extracting texture, color, and shape features from image data in these different spectral bands, multidimensional feature vectors can be constructed, thereby more accurately identifying and classifying anomaly types. For example, some anomalies may be inconspicuous under visible light but may exhibit significant differences under infrared or ultraviolet light. This multidimensional feature analysis enables the system to more robustly identify microbubbles, fiber agglomerates, or adhesive residues, and apply customized detection thresholds based on their specific types, thus avoiding misjudgments and ensuring the accuracy of subsequent noise sensitivity assessments.

[0070] In some preferred embodiments, it is assumed that the seamless underwear fabric to be detected is white polyester fiber, which may contain transparent microbubbles, white or translucent fiber clusters, and transparent or slightly yellow adhesive residues. Under visible light, these anomalies may not show a significant difference in brightness from the fabric background, or may be difficult to distinguish from each other. In this case, a multispectral imaging device can be used to acquire images of various local areas, for example, by selecting a combination of narrow-band spectral bands such as visible light (e.g., red, green, blue), near-infrared (NIR, e.g., 850 nm), and ultraviolet (UV, e.g., 365 nm). Under visible light, color and texture features can be extracted to identify anomalies that differ significantly from the fabric color. Under near-infrared light, the absorption and reflection characteristics of polyester fibers, bubbles, fiber clusters, and adhesive residues may differ significantly. For example, some adhesive residues may exhibit higher absorption rates under near-infrared light, while bubbles may exhibit different scattering characteristics. By extracting texture and shape features from the near-infrared image, these anomalies that are difficult to distinguish under visible light can be effectively differentiated. Under ultraviolet light, some organic adhesive residues may exhibit fluorescence, while fabric and bubbles do not. In such cases, adhesive residues can be specifically identified by extracting brightness or fluorescence intensity features from the ultraviolet image. By comprehensively analyzing these multispectral texture, color, and shape features—for example, using a trained support vector machine classifier—microbubbles, fiber agglomerates, or adhesive residues in localized areas can be accurately identified. For instance, if a circular area with specific scattering characteristics in the near-infrared band is detected, it is identified as a microbubble; if an irregularly shaped area exhibits fluorescence in the ultraviolet band, it is identified as adhesive residue. Based on the identified anomaly type, the system selects and applies a corresponding anomaly detection threshold. For example, for identified microbubbles, a threshold based on roundness and size might be applied; for adhesive residues, a threshold based on fluorescence intensity and area might be applied. Finally, based on these thresholds, pixel data from these anomalous areas are accurately identified and excluded, ensuring the accuracy of subsequent noise sensitivity assessments.

[0071] The steps for performing multispectral analysis on the original image data to obtain image data in different spectral bands include: Obtain information on the fabric material, color, and gloss of the seamless underwear to be tested; Based on the fabric material, color, and luster of the seamless underwear, select a narrowband spectral band combination; Under the selected combination of narrowband spectral bands, the original image data is imaged using a multispectral imaging device to obtain image data under different spectral bands; Spectral correction is performed on image data in different spectral bands.

[0072] Obtaining information on the fabric material, color, and luster of the seamless underwear to be tested refers to understanding and recording the physical and optical properties of the seamless underwear before conducting multispectral analysis. This information is crucial for subsequently selecting appropriate spectral bands, with the aim of ensuring that the selected spectral bands can highlight the spectral differences between abnormal and normal areas to the greatest extent possible.

[0073] Furthermore, selecting a narrowband spectral band combination based on the fabric material, color, and luster information of the seamless underwear involves selectively choosing several narrowband spectral bands based on pre-acquired fabric characteristics. For example, for fabrics of specific materials and colors, certain bands may be more sensitive to microbubbles or adhesive residues, while other bands may be more effective at detecting fiber aggregation. By selecting the optimal narrowband spectral band combination, the distinguishability of anomalous features can be improved, with the aim of optimizing the accuracy of subsequent image analysis.

[0074] In practical applications, using multispectral imaging equipment to image the original image data under a selected narrowband spectral band combination to obtain image data in different spectral bands refers to using specialized multispectral imaging equipment to acquire images of seamless underwear within the selected narrowband spectral bands. Multispectral imaging equipment can capture light information within different wavelength ranges, thereby generating a series of images in different spectral channels. Its purpose is to provide rich spectral information for a more comprehensive analysis of abnormal features in the images.

[0075] Furthermore, spectral correction of image data in different spectral bands refers to preprocessing the acquired multispectral image data to eliminate or reduce spectral distortion caused by factors such as uneven light sources, differences in sensor response, and environmental factors. Spectral correction can include radiometric correction, geometric correction, and atmospheric correction, with the aim of ensuring that image data in different spectral bands have accurate and comparable spectral responses, providing reliable basic data for subsequent feature extraction and anomaly identification.

[0076] refer to Figure 2 This application proposes a high-precision positioning-based image recognition system for underwear adhesive application, used to implement a high-precision positioning-based image recognition method for underwear adhesive application. The system includes: The acquisition module is used to acquire image data of seamless underwear; The calculation module is used to divide the seamless underwear image data into multiple local regions and calculate the average brightness value, brightness standard deviation and brightness gradient of each local region to obtain local brightness feature information. The determination module is used to determine the edge detection threshold of each local region based on the local brightness feature information of each local region; The extraction module is used to perform multi-resolution processing on seamless underwear image data, generate multiple multi-resolution images, apply edge detection operators to each multi-resolution image, and extract the edge map of each multi-resolution image by combining the edge detection threshold of each local region. The fusion module is used to fuse the various edge maps to obtain a preliminary edge contour. The processing module is used to perform morphological processing, smoothing, and connection processing on the initial edge contour to obtain a refined adhesive coating contour. The output module is used to convert the refined adhesive application contour into a sequence of coordinate points and output the sequence of coordinate points as an adhesive application path instruction.

[0077] Specifically, the acquisition module is configured to capture images of the seamless underwear to be inspected using an image acquisition device, such as a high-resolution industrial camera, and store and transmit these image data as seamless underwear image data.

[0078] The calculation module is configured to receive seamless underwear image data provided by the acquisition module and process the data. This processing includes dividing the image data into multiple preset local regions and, for each local region, calculating its average brightness value, brightness standard deviation, and brightness gradient. Through these calculations, local brightness feature information reflecting the illumination, texture, and edge characteristics of each local region can be obtained.

[0079] The determination module is configured to receive local brightness feature information output by the calculation module. Based on this feature information, the determination module can adaptively set an appropriate edge detection threshold for each local region to adapt to changes in illumination and texture in different regions, thereby improving the accuracy of edge detection.

[0080] The extraction module is configured to receive seamless underwear image data and perform multi-resolution processing on it to generate a series of images with different resolutions. Subsequently, the extraction module applies preset edge detection operators, such as the Canny operator or the Sobel operator, to these multi-resolution images and, in conjunction with the edge detection thresholds for each local region provided by the determination module, extracts the corresponding edge map from each multi-resolution image.

[0081] The fusion module is configured to receive the edge maps output by the extraction module. This module integrates edge maps from different resolutions using a specific fusion algorithm to compensate for potential detail loss or noise interference in single-resolution edge maps, ultimately obtaining a more complete and accurate preliminary edge contour.

[0082] The processing module is configured to receive the preliminary edge contour provided by the fusion module. To further optimize the contour quality, the processing module performs a series of post-processing operations on the preliminary edge contour, including morphological processing (such as dilation, erosion, opening, closing, etc.), smoothing processing (such as Gaussian smoothing, median filtering, etc.), and connection processing (such as connecting broken edges, removing isolated points, etc.), thereby obtaining a refined adhesive coating contour with higher accuracy and continuity.

[0083] The output module is configured to receive the refined adhesive application profile output by the processing module. This module converts the refined adhesive application profile into a series of ordered coordinate points, forming a coordinate point sequence. This coordinate point sequence is then output as the adhesive application path instruction for the adhesive application equipment, guiding the equipment to accurately complete the adhesive application operation on the underwear.

[0084] The system of this application decomposes the complex underwear adhesive image recognition method into a series of executable steps in a modular manner, which are completed collaboratively by the corresponding modules.

[0085] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A method for recognizing images of adhesive coating on underwear based on high-precision positioning, characterized in that, The method includes the following steps: Acquire image data of seamless underwear; The seamless underwear image data is divided into multiple local regions, and the average brightness value, brightness standard deviation and brightness gradient of each local region are calculated to obtain local brightness feature information. Based on the local brightness feature information of each local region, determine the edge detection threshold of each local region; Multi-resolution processing is performed on seamless underwear image data to generate multiple multi-resolution images. An edge detection operator is applied to each multi-resolution image, and the edge map of each multi-resolution image is extracted by combining the edge detection threshold of each local region. By fusing the edge images, a preliminary edge contour is obtained; The initial edge contour is subjected to morphological processing, smoothing, and connection processing to obtain a refined adhesive coating contour. The refined adhesive application contour is converted into a sequence of coordinate points, and the sequence of coordinate points is output as an adhesive application path instruction. The step of fusing the edge maps to obtain a preliminary edge contour includes: Local brightness consistency analysis is performed on each edge map. The average brightness difference between the high-resolution edge map and the low-resolution edge map in each local region is calculated. Based on the average brightness difference, it is determined whether there is distortion caused by inconsistent illumination in each local region, and the results of the local brightness consistency analysis are obtained. Calculate the local contrast intensity and connectivity of edge pixels in each edge map, and mark weak edge regions in the high-resolution edge map based on the local contrast intensity and connectivity. Based on the labeling results, strong edge regions in high-resolution edge maps are given priority. For weak edge regions in high-resolution edge maps, medium-resolution edge information is used to supplement them and make up for the lack of detail in the high-resolution edge maps. For weak edges or missing regions in the medium-resolution edge map, the overall contour information in the low-resolution edge map is weighted and corrected based on the results of local brightness consistency analysis. The corrected overall contour information in the low-resolution edge map is then filled into the medium-resolution edge map to obtain the preliminary edge contour.

2. The method for recognizing images of underwear adhesive application based on high-precision positioning as described in claim 1, characterized in that, The steps for calculating the local contrast intensity and connectivity of edge pixels in each edge map include: Texture feature analysis is performed on local regions of each edge map to obtain texture feature information; Based on texture feature information, the local contrast intensity threshold of local regions in each edge map is corrected; Calculate the local contrast intensity for edge pixels in each edge map, and determine the edge pixels based on the corrected local contrast intensity threshold; Connectivity analysis is performed on edge pixels to obtain connected edge segments; The length, rate of curvature change, and width consistency of connected edge segments are analyzed to obtain the geometric feature analysis results. Based on the geometric feature analysis results, connected edge segments that do not conform to the glue coating edge characteristics are removed, and the remaining connected edge segments are obtained. The stability of the remaining connected edge segments is evaluated, and the stability evaluation results are obtained. Based on the stability assessment results, unstable edge segments are removed.

3. The method for recognizing images of underwear adhesive application based on high-precision positioning as described in claim 1, characterized in that, The steps for performing local brightness consistency analysis on each edge map further include: Based on the average brightness difference and the known optical reflection characteristic curves of seamless underwear fabric, the average brightness difference is mapped to the edge information distortion value. Based on the degree of edge information distortion, the weight distribution of high-resolution edge maps and low-resolution edge maps during fusion is dynamically adjusted. Based on the distortion value of edge information, the local contrast intensity threshold of local regions in each edge map is corrected according to the local brightness consistency analysis.

4. The method for recognizing images of underwear adhesive application based on high-precision positioning as described in claim 1, characterized in that, The steps to obtain local brightness feature information include: Noise type identification is performed in each local area; Based on the identified noise type, select and apply the corresponding adaptive noise suppression filter to preprocess each local area; Statistical analysis is performed on the pixel grayscale values ​​of each preprocessed local region to calculate the average brightness value and brightness standard deviation of each local region. Apply edge detection operators to each preprocessed local region and calculate the brightness gradient of each local region; Local brightness characteristic information is obtained based on the average brightness value, brightness standard deviation, and brightness gradient of each local area.

5. The method for recognizing images of underwear adhesive application based on high-precision positioning as described in claim 4, characterized in that, The steps for identifying the noise type in each local area include: Texture features are extracted from each local region to obtain texture feature parameters; Based on the texture feature parameters, the noise sensitivity of each local region is evaluated, and the noise sensitivity evaluation results are obtained. Based on the noise sensitivity assessment results, the pixel grayscale histogram of each local region is adaptively segmented to obtain a segmented histogram. Perform statistical characteristic analysis on each segment histogram, and calculate the skewness, kurtosis and entropy of each segment histogram; Based on the skewness, kurtosis, and entropy of each segment histogram, and combined with the noise sensitivity assessment results, the noise type of each local region is identified as Gaussian noise, salt-and-pepper noise, or Poisson noise.

6. The method for recognizing images of underwear adhesive application based on high-precision positioning as described in claim 5, characterized in that, The steps for assessing the noise sensitivity of each local area include: Anomaly detection is performed on each local area to identify abnormal areas such as microbubbles, fiber agglomerates, or adhesive residue; When evaluating the noise sensitivity of each local region, pixel data from abnormal regions are excluded; Based on the texture feature parameters, the pixel grayscale distribution characteristics of each local region excluding abnormal areas are analyzed to obtain the masking effect parameters of the inherent texture of the fabric on noise. Based on the masking effect parameters, the noise sensitivity assessment weights for each local region are adjusted to obtain the noise sensitivity assessment results.

7. The method for recognizing images of underwear adhesive application based on high-precision positioning as described in claim 6, characterized in that, The steps for detecting abnormal points in various local areas and identifying abnormal areas such as microbubbles, fiber agglomerates, or adhesive residue include: Image data is acquired by acquiring images of each local area; Multispectral analysis was performed on the original image data to obtain image data in different spectral bands; Multispectral texture features are obtained by extracting texture features from image data in different spectral bands. Color features are extracted from image data in different spectral bands to obtain multispectral color features; Shape features are extracted from image data in different spectral bands to obtain multispectral shape features; Based on multispectral texture features, multispectral color features, and multispectral shape features, identify the anomaly types in local regions; Based on the anomaly type in the local area, select and apply the corresponding anomaly detection threshold; Based on the anomaly detection threshold, abnormal areas such as microbubbles, fiber agglomerates, or adhesive residues are identified.

8. The method for recognizing images of underwear adhesive application based on high-precision positioning as described in claim 7, characterized in that, The steps for performing multispectral analysis on the original image data to obtain image data in different spectral bands include: Obtain information on the fabric material, color, and gloss of the seamless underwear to be tested; Based on the fabric material, color, and luster of the seamless underwear, select a narrowband spectral band combination; Under the selected combination of narrowband spectral bands, the original image data is imaged using a multispectral imaging device to obtain image data under different spectral bands; Spectral correction is performed on image data in different spectral bands.

9. A high-precision positioning-based image recognition system for underwear adhesive application, applied to the high-precision positioning-based image recognition method for underwear adhesive application as described in claim 1, characterized in that, The system includes: The acquisition module acquires image data of seamless underwear; The calculation module divides the seamless underwear image data into multiple local regions and calculates the average brightness value, brightness standard deviation, and brightness gradient of each local region to obtain local brightness feature information. The determination module determines the edge detection threshold for each local region based on the local brightness feature information of each local region. The extraction module performs multi-resolution processing on the seamless underwear image data to generate multiple multi-resolution images. It then applies an edge detection operator to each multi-resolution image and, in conjunction with the edge detection thresholds for each local region, extracts the edge map of each multi-resolution image. The fusion module merges the edge images to obtain a preliminary edge contour. The processing module performs morphological processing, smoothing, and connection processing on the initial edge contour to obtain a refined adhesive coating contour. The output module converts the refined adhesive application contour into a sequence of coordinate points and outputs the sequence of coordinate points as an adhesive application path instruction. When the fusion module fuses the edge images to obtain a preliminary edge contour, it performs the following: Local brightness consistency analysis is performed on each edge map. The average brightness difference between the high-resolution edge map and the low-resolution edge map in each local region is calculated. Based on the average brightness difference, it is determined whether there is distortion caused by inconsistent illumination in each local region, and the results of the local brightness consistency analysis are obtained. Calculate the local contrast intensity and connectivity of edge pixels in each edge map, and mark weak edge regions in the high-resolution edge map based on the local contrast intensity and connectivity. Based on the labeling results, strong edge regions in high-resolution edge maps are given priority. For weak edge regions in high-resolution edge maps, medium-resolution edge information is used to supplement them and make up for the lack of detail in the high-resolution edge maps. For weak edges or missing regions in the medium-resolution edge map, the overall contour information in the low-resolution edge map is weighted and corrected based on the results of local brightness consistency analysis. The corrected overall contour information in the low-resolution edge map is then filled into the medium-resolution edge map to obtain the preliminary edge contour.

Images