Warping process monitoring method and system based on image recognition

Abstract

The invention relates to the technical field of image processing, and discloses a warping process monitoring method and system based on image recognition, and the method comprises the steps: obtaining a yarn image in the operation process of a warping machine, and inputting the image into a monitoring model; generating a direction gradient prior graph, and calculating a structure tensor representing yarn texture complexity; discrete wavelet transform is carried out on the preorder layer feature map of each dense layer, and a high-frequency detail component and a low-frequency contour component are obtained through separation; performing weighted fusion on the high-frequency detail component and the low-frequency contour component to generate a fusion feature map; performing channel splicing on the directional gradient prior image and the fusion feature image, and inputting a splicing result into a sub-pixel convolution layer for up-sampling to obtain a high-resolution monitoring image; calculating a local anisotropy index of each pixel; and the local anisotropy index and the pixel gray value are combined to identify the yarn broken end or hairiness defect, and a monitoring result is output. According to the method, the reliability of detecting tiny defects such as broken ends and hairiness under a complex background is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a method and system for monitoring the warping process based on image recognition. Background Technology

[0002] Warping is a crucial preliminary process in textile production. Its purpose is to wind a specified number of yarns onto a warp beam according to a prescribed length, width, uniform tension, and parallel arrangement. During the high-speed operation of the warping machine, the yarn is prone to defects such as breakage and fuzzing due to factors such as tension fluctuations and raw material defects.

[0003] Traditional manual inspection methods are labor-intensive, inefficient, and susceptible to subjective judgment and visual fatigue. Therefore, real-time online monitoring using machine vision technology has become an industry trend. However, existing machine vision monitoring methods still have the following drawbacks: First, traditional image processing methods (such as thresholding, edge detection, or morphological algorithms) are less robust to uneven lighting, complex backgrounds, and varied yarn textures, easily leading to missed detections and false alarms. Second, while intelligent monitoring methods based on convolutional neural networks (CNNs) can automatically extract deep features, they also face the following problems: First, the resolution of yarn images collected on-site is often limited, causing features of subtle defects such as fuzz to be easily lost during multiple rounds of downsampling in the network, thus affecting the model's recognition accuracy. Second, if general super-resolution techniques are used to improve image quality, the reconstruction network may not effectively preserve the inherent directional texture of the yarn when the image is enlarged, and may even introduce blurring or artifacts. Finally, in the defect discrimination stage, relying solely on the classifier's output results results in low reliability for defects with diverse shapes and high similarity to the background texture. Summary of the Invention

[0004] This invention provides an image recognition-based method and system for monitoring the warping process to solve the problems of missed detection and false alarms in the prior art when facing scenarios with uneven lighting, complex backgrounds and varied yarn textures.

[0005] In a first aspect, the image recognition-based warping process monitoring method of the present invention includes the following steps: Acquire yarn images during the operation of the warping machine and input the yarn images into a preset monitoring model; calculate the directional gradient of the yarn images to generate a directional gradient prior map, and calculate the structure tensor representing the yarn texture complexity based on the directional gradient prior map; In the dense connection module of the monitoring model, discrete wavelet transform is performed on the feature maps of the preceding layers of each dense layer to separate high-frequency detail components and low-frequency contour components; the structural tensor is used as prior information to guide the weighted fusion of high-frequency detail components and low-frequency contour components to generate a fused feature map containing multi-scale features. In the super-resolution reconstruction module of the monitoring model, the prior map of directional gradients is used as guiding information and is channel-stitched with the fusion feature map generated by the dense connection module. The stitching result is then input into the sub-pixel convolutional layer for upsampling to reconstruct a high-resolution monitoring image. For the high-resolution monitoring image, the local second-order moment matrix of each pixel neighborhood is calculated and the eigenvalues ​​are solved to obtain the local anisotropy index; and the yarn breakage or fuzz defects are identified by combining the local anisotropy index and the pixel gray value, and the monitoring results are output.

[0006] Preferably, the calculation of the structure tensor representing the yarn texture complexity based on the directional gradient prior map includes: For each pixel in the directional gradient prior map, select a pixel centered at the given pixel. Neighborhood window; Calculate the gradient components of all pixels within the neighborhood window in the x and y directions; A second-order moment matrix is ​​constructed based on the gradient components, and the second-order moment matrix is ​​used as the structure tensor at the pixel.

[0007] Preferably, the step of performing discrete wavelet transform on the feature maps of the preceding layers of each dense layer to separate high-frequency detail components and low-frequency contour components includes: Perform a single-layer discrete wavelet transform on the feature map of the preceding layer; The horizontal detail components, vertical detail components, and diagonal detail components obtained after transformation are collectively used as the high-frequency detail components; The approximate component obtained after transformation is taken as the low-frequency contour component.

[0008] Preferably, the step of using the structural tensor as prior information to guide the weighted fusion of high-frequency detail components and low-frequency contour components includes: Calculate two eigenvalues ​​of the structure tensor and use the difference between the two eigenvalues ​​as the texture intensity factor to form a texture intensity factor map; The texture intensity factor map is adjusted to the same spatial size as the high-frequency detail component and the low-frequency contour component through a downsampling operation; By using the Sigmoid activation function, the resized texture intensity factor map is mapped to a high-frequency weight coefficient map ranging from 0 to 1; The low-frequency weighting coefficient map is obtained by subtracting the high-frequency weighting coefficient map from 1; The high-frequency detail component is multiplied element-wise with the high-frequency weight coefficient map, and the low-frequency contour component is multiplied element-wise with the low-frequency weight coefficient map. The results of the element-wise multiplications are then added together to complete the weighted fusion.

[0009] Preferably, the step of using the directional gradient prior map as guiding information to perform channel-wise concatenation with the fusion feature map generated by the dense connection module, and inputting the concatenation result into a sub-pixel convolutional layer for upsampling to reconstruct a high-resolution monitoring image includes: The directional gradient prior map is adjusted to the same spatial size as the fused feature map through a downsampling operation; In the channel dimension, the resized orientation gradient prior map is concatenated with the fused feature map; the concatenated feature map is then input into the sub-pixel convolutional layer to reconstruct the high-resolution monitoring image.

[0010] Preferably, the step of calculating the local second-order moment matrix of each pixel neighborhood and solving for the eigenvalues ​​to obtain the local anisotropy index for the high-resolution monitoring image includes: For each pixel in the high-resolution monitoring image, construct Neighborhood window; The covariance matrix of the image gradient within the neighborhood window is calculated as the local second-order moment matrix; Solve for the two eigenvalues ​​of the local second-order moment matrix. and ,in ; The local anisotropy index is calculated using the following formula. ; ;in, This is used to prevent small positive numbers with a denominator of zero.

[0011] Preferably, the step of combining the local anisotropy index and pixel grayscale values ​​to identify yarn breakage or fuzz defects includes: When the local anisotropy index of a pixel is greater than the first anisotropy index threshold At that time, it was identified as a point of breakage defect; When the local anisotropy index of a pixel is greater than the second anisotropy index threshold And less than or equal to the first anisotropy index threshold And the grayscale value is lower than the grayscale threshold. At that time, it was identified as a defect point in the feathers; among them, .

[0012] Preferably, acquiring yarn images during the operation of the warping machine includes: An industrial camera is used to continuously acquire yarn images at a preset frame rate.

[0013] Preferably, the Haar wavelet basis is used to perform a single-layer discrete wavelet transform on the feature map of the preceding layer.

[0014] Secondly, the image recognition-based warping process monitoring system of the present invention includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the above-mentioned image recognition-based warping process monitoring method.

[0015] The beneficial effects of this invention are as follows: By calculating the directional gradient prior map and the structural tensor, this invention inputs the inherent orientation and texture complexity information of the yarn into the model. In the feature fusion stage, discrete wavelet transform is used to decompose the features into high-frequency details and low-frequency contours, and the structural tensor is used to achieve weighted fusion of the two, thereby realizing the distinction and focus on yarn contour information and texture details, enhancing the model's ability to represent complex textures. In the super-resolution reconstruction process, the directional gradient prior map is used to guide the upsampling process, which can reconstruct high-resolution images with clear texture edges and rich details, suppressing artifacts and blurring during image magnification, and making the details of small fuzzy defects clearer. Combining the local anisotropy index and pixel grayscale for defect identification provides a clear physical representation basis for defect discrimination, improving the reliability of detecting small defects such as broken ends and fuzzy textures in complex backgrounds. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the image recognition-based warping process monitoring method provided in an embodiment of the present invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0018] like Figure 1 As shown, an embodiment of the image recognition-based warping process monitoring method provided by the present invention includes the following steps: S1. Acquire yarn images during the operation of the warping machine and input the yarn images into the preset monitoring model; calculate the directional gradient of the yarn images to generate a directional gradient prior map, and calculate the structure tensor representing the yarn texture complexity based on the directional gradient prior map.

[0019] Specifically, an industrial camera is installed directly above the yarn sheet area of ​​the warping machine, equipped with a linear LED light source to provide stable and uniform illumination. The camera continuously acquires yarn images at a preset frame rate and transmits the detected digital images to an industrial control computer in real time via a data interface. A pre-trained monitoring model is deployed on the industrial control computer. This model receives the digital image as an input tensor for subsequent processing. Using the Sobel operator, the first-order partial derivatives of the input yarn image in the horizontal x-direction and the vertical y-direction are calculated to obtain the gradient map in the x-direction. and gradient plot in the y direction .Will and The values ​​at each pixel are combined to form a gradient vector, thus constructing the directional gradient prior map. For each pixel in the directional gradient prior map, a neighborhood window centered on that pixel is defined, for example, with a size of... Within this window, the sum of squares and the sum of products of the gradient components are calculated, i.e., the calculation is performed. as well as This is how the pixel is constructed. The structure tensor matrix has the following elements: The eigenvalues ​​of the structure tensor matrix reflect the texture intensity and directionality within the neighborhood, thus representing the texture complexity of the yarn.

[0020] S2, in the dense connection module of the monitoring model, discrete wavelet transform is performed on the feature maps of the preceding layers of each dense layer to separate high-frequency detail components and low-frequency contour components; the structure tensor is used as prior information to guide the weighted fusion of high-frequency detail components and low-frequency contour components to generate a fused feature map containing multi-scale features.

[0021] Specifically, for any layer in the densely connected module, after concatenating the feature maps output from all preceding layers, a two-dimensional discrete wavelet transform, such as the Haar wavelet transform, is applied to the concatenated feature map to decompose it into a low-frequency approximation component (low-frequency contour component) and high-frequency detail components in the horizontal, vertical, and diagonal directions. The calculated structure tensor map is resized to match the feature map size. A weight map is generated based on the structure tensor values, assigning higher weights to high-frequency components in areas with complex textures and higher weights to low-frequency components in areas with smooth textures. This weight map is then weighted and summed pixel-by-pixel with both the high-frequency detail component and the low-frequency contour component to generate a fused feature map that contains both clear contours and rich texture details.

[0022] S3, in the super-resolution reconstruction module of the monitoring model, the prior map of directional gradient is used as guiding information and is channel-stitched with the fusion feature map generated by the dense connection module. The stitching result is then input into the sub-pixel convolutional layer for upsampling to reconstruct a high-resolution monitoring image.

[0023] Specifically, the generated directional gradient prior map undergoes downsampling and other operations to ensure its spatial resolution matches that of the fused feature map output by the densely connected module. In the channel dimension, the directional gradient prior map is concatenated with this fused feature map to form a multi-channel feature tensor containing richer edge guidance information. This concatenated multi-channel feature tensor is then fed into a sub-pixel convolutional layer. This layer uses a standard convolution operation to increase the number of channels in the input feature map by a factor of S squared, where S is the desired upsampling factor. Through periodic rearrangement operations, the low-resolution feature map with a large number of channels is rearranged into a high-resolution feature map with a reduced number of channels but an increased spatial resolution by a factor of S. This is then output as the reconstructed high-resolution monitoring image.

[0024] S4. For the high-resolution monitoring image, calculate the local second-order moment matrix of each pixel neighborhood and solve for the eigenvalues ​​to obtain the local anisotropy index; and combine the local anisotropy index and pixel gray value to identify yarn breakage or fuzz defects and output the monitoring results.

[0025] Specifically, for each pixel of the high-resolution monitoring image, a neighborhood window is taken centered on that pixel. The gradients in the x and y directions of the image within this window are calculated, and the local second-order moment matrix of that pixel is constructed using these gradients. The two eigenvalues ​​of this matrix are then solved. and The local anisotropy index is calculated based on two eigenvalues. The formula for calculating the local anisotropy index is as follows: ,in For larger feature values, an anisotropy index threshold and a grayscale value threshold are set. When the local anisotropy index and grayscale value of a pixel are both greater than their respective thresholds, the pixel is determined to be a fuzzy defect point. For yarn breakage defects, they are identified by detecting large areas of low anisotropy index that appear consecutively along the yarn path. Based on the identified defect type and location, monitoring results are generated and output.

[0026] In an optional embodiment, the calculation of the structure tensor representing the yarn texture complexity based on the directional gradient prior map includes: for each pixel in the directional gradient prior map, selecting a structure centered on the pixel. Neighborhood window; calculate the gradient components of all pixels within the neighborhood window in the x and y directions; construct a second-order moment matrix based on the gradient components, and use the second-order moment matrix as the structure tensor at the pixel.

[0027] Specifically, each pixel in the directional gradient prior map is expanded. For example, for the pixel with coordinates 100, 50, a coordinate system centered on it is selected. A neighborhood window of size is used, containing nine pixels. The gradient components of each of these nine pixels in the horizontal (x) and vertical (y) directions are calculated, resulting in nine sets of gradient data pairs, such as... It is 1.2 and The value is -0.5. Based on these nine sets of gradient data, a gradient is constructed for the center pixel. The second-order moment matrix, i.e., the structure tensor. The top-left element of this matrix represents all pixels within the window. The sum of the squares of the components, the bottom right element is all The sum of the squares of the components, while the two diagonal elements are all and The sum of the products. Suppose the calculated sums are 25.4, 18.1, and -5.7, respectively. These values ​​represent the texture complexity and directionality of the local region.

[0028] In an optional embodiment, performing discrete wavelet transform on the feature maps of the preceding layers of each dense layer to separate high-frequency detail components and low-frequency contour components includes: A single-layer discrete wavelet transform is performed on the feature map of the preceding layer using the Haar wavelet basis; the horizontal detail components, vertical detail components, and diagonal detail components obtained after the transform are used together as the high-frequency detail components; and the approximate components obtained after the transform are used as the low-frequency contour components.

[0029] Specifically, the feature map of the preceding layer of a dense layer is used as input, for example, a feature map of size [missing information]. The single-channel feature map is obtained. A single-level discrete wavelet transform is performed on the single-channel feature map using the Haar wavelet basis. The Haar wavelet decomposes the feature map by applying low-pass and high-pass filters to the rows and columns of the image, respectively. After the transform, the original... The single-channel feature map is decomposed into four dimensions, each with a size of 1. The feature map is divided into three sub-maps. The first sub-map is the approximate component, which is essentially a scaled-down version of the original feature map, preserving the main contours and low-frequency information. The other three sub-maps are the horizontal detail component, the vertical detail component, and the diagonal detail component, which detect high-frequency information such as edges and textures in the corresponding directions of the original features, respectively. These three detail components are combined to form the high-frequency detail component, while the approximate component is treated separately as the low-frequency contour component, thus achieving feature separation.

[0030] In an optional embodiment, using the structural tensor as prior information to guide the weighted fusion of high-frequency detail components and low-frequency contour components includes: calculating two eigenvalues ​​of the structural tensor and using the difference between the two eigenvalues ​​as a texture intensity factor to form a texture intensity factor map; adjusting the texture intensity factor map to the same spatial size as the high-frequency detail components and low-frequency contour components through a downsampling operation; mapping the adjusted texture intensity factor map to a high-frequency weight coefficient map ranging from 0 to 1 using a sigmoid activation function; obtaining a low-frequency weight coefficient map by subtracting the high-frequency weight coefficient map from 1; performing element-wise multiplication of the high-frequency detail components with the high-frequency weight coefficient map, performing element-wise multiplication of the low-frequency contour components with the low-frequency weight coefficient map, and adding the element-wise multiplication results to complete the weighted fusion.

[0031] Specifically, for each pixel in the structure tensor graph... The structure tensor matrix is ​​used to calculate its two eigenvalues. The difference between these two eigenvalues ​​is used as the texture intensity factor for that point; a larger difference indicates a more prominent texture. For example, an eigenvalue difference of 50.3 indicates a strong yarn texture at that point, while a difference of 1.2 indicates a smoother area. By iterating through all pixels and performing the above calculations, a texture intensity factor map with the same spatial dimensions as the original image is finally generated.

[0032] To match the spatial dimensions of the texture intensity factor map with the feature components after wavelet transform, a downsampling operation needs to be performed on the texture intensity factor map, reducing its size from... Shrink to Next, the resized texture intensity factor map is input into a sigmoid activation function, which maps the texture intensity factor value of each pixel in the map to a range of 0 to 1, generating a high-frequency weight coefficient map. For example, a factor with a value of 50.3 is mapped to 0.99, while a factor with a value of 1.2 is mapped to 0.55. Finally, the corresponding low-frequency weight coefficient map is obtained by subtracting the high-frequency weight coefficient map from the constant 1.

[0033] The high-frequency detail component is multiplied element-wise with the high-frequency weight coefficient map, and the low-frequency contour component is multiplied element-wise with the low-frequency weight coefficient map. The two products are then added together. In regions with complex textures, high-frequency details are enhanced by a weight close to 1; in smooth regions, low-frequency contours are preserved by a higher weight, thus achieving feature fusion guided by yarn texture complexity.

[0034] In an optional embodiment, the step of using the directional gradient prior map as guiding information to perform channel concatenation with the fusion feature map generated by the dense connection module, and inputting the concatenation result into a sub-pixel convolutional layer for upsampling to reconstruct a high-resolution monitoring image includes: The directional gradient prior map is downsampled to the same spatial size as the fused feature map; in the channel dimension, the resized directional gradient prior map is concatenated with the fused feature map; the concatenated feature map is input into the subpixel convolutional layer to reconstruct the high-resolution monitoring image.

[0035] For example, the super-resolution reconstruction module needs to process two input data: one is a data set with a size of [missing information]. The first is the directional gradient prior map, and the second is the fused feature map generated by the densely connected modules. Assuming the size of the fused feature map is... It has 32 channels. To standardize the size, it is necessary to... The directional gradient prior map is resized by downsampling operations, for example, using bilinear interpolation. After standardizing the dimensions, the single-channel... Oriented gradient prior map and 32-channel The fused feature maps are stitched together along the channel dimension to form a shape of size [size missing]. It also features a new feature map with 33 channels. In this way, clear yarn edges and texture direction information are directly incorporated into the abstract features extracted by the deep learning network.

[0036] The combined feature map containing 33 channels is fed into a sub-pixel convolutional layer. Subsequently, this combined feature map containing 33 channels is fed into the sub-pixel convolutional layer for processing. The core function of this layer is to increase the resolution of the feature map, for example, by reducing its resolution from... Zoom in Through specific convolution and periodic rearrangement operations, subpixel convolutional layers utilize learned features and infused gradient-guided information to meticulously reconstruct high-resolution monitoring images, thereby ensuring rich details and clear edges in the reconstruction results.

[0037] In an optional embodiment, the step of calculating the local second-order moment matrix of each pixel neighborhood and solving for the eigenvalues ​​to obtain the local anisotropy index for the high-resolution monitoring image includes: constructing for each pixel in the high-resolution monitoring image... Neighborhood window; calculate the covariance matrix of the image gradient within the neighborhood window as the local second-order moment matrix; solve for the two eigenvalues ​​of the local second-order moment matrix. and ,in The local anisotropy index is calculated using the following formula. ; ;in, This is used to prevent small positive numbers with a denominator of zero.

[0038] Specifically, each pixel in the generated high-resolution monitoring image is processed. Taking any pixel in the image as an example, a [process] is selected centered on that pixel. A neighborhood window of size 25 pixels is generated. The image gradient of each of these 25 pixels is then calculated, where each gradient consists of a horizontal component and a vertical component. Using these 25 gradient vectors, their covariance matrix is ​​calculated. The covariance matrix is ​​the local second-order moment matrix of the central pixel, representing the overall trend and dispersion of the gradient distribution in that neighborhood. The two eigenvalues ​​of this local second-order moment matrix are denoted as [equations to be found here]. and The magnitude and relative difference of these two eigenvalues ​​reflect the directionality of the local texture. For example, if For version 15.0 A value of 0.5 indicates that the texture is very significant in a specific direction, but extremely insignificant in the direction perpendicular to it. Then, the local anisotropy index for this pixel is calculated using the formula to be 0.935. A local anisotropy index close to 1 indicates that the region containing this pixel has extremely strong directionality, such as obvious straight edges or fibrous textures. By repeating the above calculation process for all pixels, a complete local anisotropy index map can be generated.

[0039] In an optional embodiment, the step of combining the local anisotropy index and pixel grayscale value to identify yarn breakage or fuzz defects includes: when the local anisotropy index of a pixel is greater than a first anisotropy index threshold... At that time, it was identified as a point of breakage defect; When the local anisotropy index of a pixel is greater than the second anisotropy index threshold And less than or equal to the first anisotropy index threshold And the grayscale value is lower than the grayscale threshold. At that time, it was identified as a defect point in the feathers; among them, .

[0040] Specifically, based on empirical data, a first anisotropy index threshold is set. The threshold value is 0.9, representing the second anisotropy index. The grayscale threshold is 0.6. It is 50. Used to detect structural defects with extremely strong decapitation structures, This is used to define defects in feather-like structures with slightly weaker directionality; grayscale threshold. This is used to distinguish darker defect areas from brighter yarn bodies. During detection, the calculated local anisotropy index and the grayscale value of the pixel itself are acquired. For a pixel, if its local anisotropy index is 0.95, it is considered a positive result because its local anisotropy index is greater than the set threshold. This pixel was identified as a breakage defect. This is because a break in the yarn creates a highly directional and abrupt edge, resulting in an extremely high local anisotropy index.

[0041] If the local anisotropy index of a pixel is 0.7, since 0.7 is greater than 0.7, the anisotropy index of a pixel is 0.7. 0.6 and not greater than If the value is 0.9, then check its grayscale value. Assume the grayscale value of this pixel is 40, which is lower than the set grayscale threshold. If a pixel is found to have a fuzzy texture, it is considered a defective pixel. This combined discrimination logic can effectively identify pixels that have a certain directionality but a grayscale value lower than that of normal yarn areas, which is a typical characteristic of fuzzy texture defects. Pixels that do not meet any of the above defect discrimination conditions are considered normal.

[0042] The implementation principle of the image recognition-based warping process monitoring method of this invention is as follows: This invention explicitly introduces the inherent directional information and texture complexity information of the yarn into the monitoring model by calculating the directional gradient prior map and the structural tensor. In the feature fusion stage, the feature map is decomposed into high-frequency detail components and low-frequency contour components using discrete wavelet transform; and the structural tensor guides the weighted fusion of the two. This design achieves differentiated processing and focused attention on yarn contour information and texture details, thereby enhancing the model's ability to represent complex textures. In the super-resolution reconstruction process, the directional gradient prior map guides the upsampling process. This helps to reconstruct a high-resolution image with clear texture edges and rich details, effectively suppressing artifacts and blurring during image magnification, and allowing the details of minute fuzz defects to be clearly presented. Finally, in the defect identification stage, this invention combines the local anisotropy index and pixel grayscale values ​​for comprehensive discrimination. This provides a clear physical basis for defect identification and improves the reliability of detecting minute defects such as broken ends and fuzz in complex backgrounds.

[0043] An embodiment of the image recognition-based warping process monitoring system provided by the present invention includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the image recognition-based warping process monitoring method in the above embodiment.

[0044] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for monitoring the warping process based on image recognition, characterized in that, The process includes the following steps: acquiring yarn images during the warping machine's operation and inputting the yarn images into a preset monitoring model; calculating the directional gradient of the yarn images to generate a directional gradient prior map, and calculating the structure tensor representing the yarn texture complexity based on the directional gradient prior map; In the dense connection module of the monitoring model, discrete wavelet transform is performed on the feature maps of the preceding layers of each dense layer to separate high-frequency detail components and low-frequency contour components; the structural tensor is used as prior information to guide the weighted fusion of high-frequency detail components and low-frequency contour components to generate a fused feature map containing multi-scale features. In the super-resolution reconstruction module of the monitoring model, the prior map of directional gradients is used as guiding information and is channel-stitched with the fusion feature map generated by the dense connection module. The stitching result is then input into the sub-pixel convolutional layer for upsampling to reconstruct a high-resolution monitoring image. For the high-resolution monitoring image, the local second-order moment matrix of each pixel neighborhood is calculated and the eigenvalues ​​are solved to obtain the local anisotropy index; and the yarn breakage or fuzz defects are identified by combining the local anisotropy index and the pixel gray value, and the monitoring results are output.

2. The method for monitoring the warping process based on image recognition according to claim 1, characterized in that, The calculation of the structure tensor representing yarn texture complexity based on the directional gradient prior map includes: For each pixel in the directional gradient prior map, select a pixel centered at the given pixel. Neighborhood window; Calculate the gradient components of all pixels within the neighborhood window in the x and y directions; A second-order moment matrix is ​​constructed based on the gradient components, and the second-order moment matrix is ​​used as the structure tensor at the pixel.

3. The method for monitoring the warping process based on image recognition according to claim 1, characterized in that, The discrete wavelet transform of the feature maps of the preceding layers in each dense layer is performed to separate high-frequency detail components and low-frequency contour components, including: Perform a single-layer discrete wavelet transform on the feature map of the preceding layer; The horizontal detail components, vertical detail components, and diagonal detail components obtained after transformation are collectively used as the high-frequency detail components; The approximate component obtained after transformation is taken as the low-frequency contour component.

4. The method for monitoring the warping process based on image recognition according to claim 1, characterized in that, The step of using the structural tensor as prior information to guide the weighted fusion of high-frequency detail components and low-frequency contour components includes: Calculate two eigenvalues ​​of the structure tensor and use the difference between the two eigenvalues ​​as the texture intensity factor to form a texture intensity factor map; The texture intensity factor map is adjusted to the same spatial size as the high-frequency detail component and the low-frequency contour component through a downsampling operation; By using the Sigmoid activation function, the resized texture intensity factor map is mapped to a high-frequency weight coefficient map ranging from 0 to 1; The low-frequency weighting coefficient map is obtained by subtracting the high-frequency weighting coefficient map from 1; The high-frequency detail component is multiplied element-wise with the high-frequency weight coefficient map, and the low-frequency contour component is multiplied element-wise with the low-frequency weight coefficient map. The results of the element-wise multiplications are then added together to complete the weighted fusion.

5. The method for monitoring the warping process based on image recognition according to claim 1, characterized in that, The process involves using the directional gradient prior map as guiding information, performing channel-wise concatenation with the fusion feature map generated by the dense connection module, and inputting the concatenation result into a sub-pixel convolutional layer for upsampling to reconstruct a high-resolution monitoring image, including: The directional gradient prior map is adjusted to the same spatial size as the fused feature map through a downsampling operation; In the channel dimension, the resized orientation gradient prior map is concatenated with the fused feature map; the concatenated feature map is then input into the sub-pixel convolutional layer to reconstruct the high-resolution monitoring image.

6. The method for monitoring the warping process based on image recognition according to claim 1, characterized in that, The process of calculating the local second-order moment matrix of each pixel neighborhood and solving for the eigenvalues ​​in the high-resolution monitoring image to obtain the local anisotropy index includes: For each pixel in the high-resolution monitoring image, construct Neighborhood window; The covariance matrix of the image gradient within the neighborhood window is calculated as the local second-order moment matrix; Solve for the two eigenvalues ​​of the local second-order moment matrix. and ,in ; The local anisotropy index is calculated using the following formula. ; ;in, This is used to prevent small positive numbers with a denominator of zero.

7. The method for monitoring the warping process based on image recognition according to claim 1, characterized in that, The method of identifying yarn breakage or fuzziness defects by combining local anisotropy index and pixel grayscale values ​​includes: When the local anisotropy index of a pixel is greater than the first anisotropy index threshold At that time, it was identified as a point of breakage defect; When the local anisotropy index of a pixel is greater than the second anisotropy index threshold And less than or equal to the first anisotropy index threshold And the grayscale value is lower than the grayscale threshold. At that time, it was identified as a defect point in the feathers; among them, .

8. The method for monitoring the warping process based on image recognition according to claim 1, characterized in that, The acquisition of yarn images during the warping machine operation includes: An industrial camera is used to continuously acquire yarn images at a preset frame rate.

9. The method for monitoring the warping process based on image recognition according to claim 3, characterized in that, A single-layer discrete wavelet transform is performed on the feature map of the preceding layer using the Haar wavelet basis.

10. A warping process monitoring system based on image recognition, characterized in that, It includes a memory and a processor. The memory stores computer instructions. When the processor executes the computer instructions, it implements the image recognition-based warping process monitoring method as described in any one of claims 1-9.

Images