Defect intelligent identification method and system for spunlace nonwoven fabric

Abstract

The application relates to the field of image processing, in particular to a spunlace non-woven fabric defect intelligent identification method and system, which comprises the following steps: dividing a standard non-defect spunlace non-woven fabric image into multiple local image blocks according to spatial positions; constructing a local spectrum prior atlas corresponding to the image spatial positions; performing block processing on a to-be-detected image to obtain multiple original image blocks, and dynamically generating a personalized frequency domain mask for each original image block based on the local spectrum prior atlas to suppress the frequency band corresponding to normal textures; constructing a preliminary defect residual graph and a spectrum graph thereof; in response to the proportion of the medium-low frequency region energy in the total frequency domain energy in the spectrum graph exceeding a preset threshold, updating the mask parameters, applying the updated mask parameters to the personalized frequency domain mask, and obtaining a final defect residual graph; performing threshold segmentation and / or connected domain analysis on the final defect residual graph, and outputting a defect position. The application has the effect of improving defect detection accuracy.

Description

Technical Field

[0001] This application relates to the field of image processing, and in particular to a method and system for intelligent defect recognition of spunlace nonwoven fabric. Background Technology

[0002] Spunlace nonwoven fabric is a nonwoven material formed by spraying high-pressure micro-jet water onto one or more layers of fiber web, causing the fibers to entangle and reinforce each other. Due to its advantages such as softness, breathability, non-toxicity, and biodegradability, it is widely used in medical and health care, wiping materials, clothing linings, and household products.

[0003] As downstream applications increasingly demand higher standards for product appearance and performance, spunlace nonwoven fabrics may encounter various defects during production, such as holes, stains, uneven thickness, fiber agglomeration, and edge burrs, severely impacting product quality and customer experience. In recent years, with the development of machine vision, deep learning, and industrial automation technologies, intelligent defect recognition methods based on image processing have gradually become a research hotspot.

[0004] During the production of spunlace nonwoven fabric, due to factors such as fixed spunlace direction, uneven tension distribution, and random fiber distribution, the normal fabric surface itself has different texture characteristics in different spatial areas. However, the inherent texture of the fabric surface and the irregular patterns formed by the spunlace process are easily misjudged as defects, resulting in a high false alarm rate. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method and system for intelligent defect identification of spunlace nonwoven fabrics.

[0006] In a first aspect, this application provides a method for intelligent defect identification of spunlace nonwoven fabrics, employing the following technical solution:

[0007] A method for intelligent defect identification in spunlace nonwoven fabrics, comprising the following steps:

[0008] The acquired standard defect-free spunlace nonwoven fabric image is divided into multiple local image blocks according to spatial location. A local spectral prior map is constructed, corresponding one-to-one with the spatial location of the image. Each local image block is spatially aligned according to a preset repetition period of the spunlace nonwoven fabric surface texture. The image to be inspected is processed into multiple original image blocks. Based on the local spectral prior map, a personalized frequency domain mask is dynamically generated for each original image block to suppress the frequency band corresponding to the normal texture. The personalized frequency domain mask is applied to the frequency domain representation of the original image block, and a preliminary defect residual map is obtained through amplitude-frequency transformation. The preliminary defect residual map is then subjected to frequency-amplitude transformation to obtain a spectrum map. In response to the proportion of energy in the mid-to-low frequency region of the spectrum map exceeding a preset threshold, the mask parameters are updated. The updated mask parameters are applied to the personalized frequency domain mask to obtain the final defect residual map. The mask parameters include the frequency band corresponding to the normal texture and the suppression intensity applied to that frequency band to suppress the normal texture component. Threshold segmentation and / or connected component analysis are performed on the final defect residual map to output the defect location.

[0009] The mask parameter update method is as follows: calculate the local unsuppressed responsivity corresponding to the current suppression intensity to characterize the degree of residue of normal texture components in the residual spectrum; calculate the optimal adjustment step size based on the local unsuppressed responsivity of the current preliminary defect residual map and the average local unsuppressed responsivity of the preliminary defect residual maps in its neighborhood, and adjust the suppression intensity according to the optimal adjustment step size; calculate the noise level, adjust the preset initial frequency band according to the noise level to obtain the optimal frequency band; update the mask parameters according to the optimal frequency band and the suppression intensity.

[0010] Optionally, the method for calculating the local unsuppressed response includes the following steps: under the current mask parameters, obtain the energy ratio of low frequency and full frequency domain in the original image patch as a reference value; apply a perturbation amount to the suppression intensity to generate a perturbation mask; process the original image patch using the perturbation mask to obtain a residual image; and calculate the ratio of the energy ratio of low frequency and full frequency domain in the residual image to the reference value as the local unsuppressed response of the suppression intensity.

[0011] Optionally, the method for calculating the optimal adjustment step size includes the following steps: setting an initial step size; calculating the relative unsuppressed responsivity; in response to the relative unsuppressed responsivity being greater than a preset responsivity threshold, the product of the initial step size and the relative unsuppressed responsivity is taken as the optimal adjustment step size, otherwise the optimal adjustment step size is 0; wherein, the method for calculating the relative unsuppressed responsivity is: normalizing the absolute difference between the local unsuppressed responsivity and the mean of the neighborhood unsuppressed responsivity and taking it as the relative unsuppressed responsivity; or normalizing the ratio between the local unsuppressed responsivity and the mean of the neighborhood unsuppressed responsivity and taking it as the relative unsuppressed responsivity.

[0012] Optionally, the response threshold can be set by: calculating the standard deviation of the local unsuppressed response of each image block in the neighborhood of the current image block; and setting the response threshold as the product of the standard deviation and a preset multiplier.

[0013] Optionally, constructing a local spectral prior map that corresponds one-to-one with the spatial location of the image includes: performing a two-dimensional Fourier transform on each local image patch and extracting local spectral features to construct a local spectral prior map.

[0014] Optionally, local spectral features include the dominant frequency direction, energy concentration bandwidth, and frequency domain energy entropy.

[0015] Optionally, in the threshold segmentation of the final defect residual map, the threshold can be an adaptive threshold or a fixed threshold.

[0016] Optionally, the noise level is the signal-to-noise ratio.

[0017] Secondly, this application provides a defect intelligent identification system for spunlace nonwoven fabrics, which adopts the following technical solution:

[0018] A defect intelligent identification system for spunlace nonwoven fabric includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the defect intelligent identification method for spunlace nonwoven fabric described above is implemented.

[0019] The beneficial effect is that the above-mentioned intelligent defect identification method for spunlace nonwoven fabric is used to generate a computer program and stored in the memory so that it can be loaded and executed by the processor. Thus, a system can be made based on the memory and the processor, which is convenient to use.

[0020] This application has the following technical advantages:

[0021] 1. Due to the manufacturing process, spunlace nonwoven fabrics exhibit regular or random textures on their surface. These normal textures may have similar frequency domain characteristics to defect areas in images, easily leading to misjudgment as defects. Local spectral prior maps can accurately characterize the distribution characteristics of normal textures on the spunlace nonwoven fabric surface in the frequency domain. The characteristics of normal textures in different regions may vary significantly, and a fixed suppression intensity cannot meet the needs of all scenarios. By calculating the optimal adjustment step size, the system can dynamically adjust the suppression intensity based on the local unsuppressed responsivity of the current region and its neighborhood consistency, thereby achieving a more accurate background suppression effect.

[0022] 2. Mask parameters include the frequency band corresponding to the normal texture and the suppression intensity applied to that frequency band to suppress the normal texture component; the signal-to-noise ratio reflects the image quality, the frequency band corresponding to the normal texture defines the frequency domain range that needs to be suppressed, and the suppression intensity reflects how much suppression is needed in the current region to fully weaken the influence of the normal texture.

[0023] 3. The texture formed on the surface of spunlace nonwoven fabric due to the manufacturing process has a certain periodicity or regularity. Through physical alignment, it can be ensured that each local image block fully represents the local statistical characteristics of the normal texture in that area. After alignment, local image blocks at the same location ideally have similar frequency domain feature distributions, ensuring that the mask parameters can accurately correspond to the corresponding areas of the image to be detected, thereby improving the generation accuracy of personalized frequency domain masks. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for intelligent defect identification of spunlace nonwoven fabric according to an embodiment of this application.

[0025] Figure 2 This is a flowchart of step S3 in a method for intelligent defect identification of spunlace nonwoven fabric according to an embodiment of this application. Detailed Implementation

[0026] This application discloses a method for intelligent defect identification in spunlace nonwoven fabrics, referring to... Figure 1 The process includes steps S1-S4, as detailed below:

[0027] S1: Divide the collected standard defect-free spunlace nonwoven fabric images into multiple local image blocks according to their spatial location; construct a local spectral prior map that corresponds one-to-one with the spatial location of the images.

[0028] Specifically, the standard defect-free spunlace nonwoven fabric image serves as a reference image for characterizing normal texture characteristics. Each local image patch is spatially aligned according to a preset repetition period of the spunlace nonwoven fabric surface texture. Based on this preset repetition period, the spatial domain is divided into multiple local image patches. Image segmentation can be performed using a custom grid, such as a fixed-size regular grid (e.g., 64×64, 128×128 pixels) for uniform segmentation. This ensures that each local image patch fully represents the local statistical characteristics of the normal texture in that region and, ideally, exhibits a similar frequency domain feature distribution. Alternatively, the image can be based on a main direction, with repetition intervals at integer multiples along that main direction. (like , Set the block boundaries, and also set appropriate dimensions (which can be) in the direction perpendicular to the main direction. Divide the image into rectangular or parallelogram local image patches (or multiples thereof, or adaptively based on the image height and width) to generate a set of local image patches with boundaries aligned with the texture period.

[0029] After the segmentation is completed, a two-dimensional Fourier transform is performed on each local image block to convert it from the spatial domain to the frequency domain, thus obtaining the corresponding complex spectrum representation.

[0030] Local spectral features include the dominant frequency direction, energy concentration bandwidth, and frequency domain energy entropy. Specifically, the dominant frequency direction can be determined by detecting the position of energy peaks in the spectral amplitude plot and calculating their angle relative to the spectral center (i.e., the DC component), which serves as the dominant direction of the local block texture. Accurate estimation can be achieved using Hough transform, polar coordinate projection, or gradient orientation histograms (such as Radon transform). The energy concentration bandwidth is defined as the width of the smallest annular or sector-shaped region containing a predetermined proportion (e.g., 90%) of the frequency domain energy, used to characterize the degree of energy concentration in the frequency domain for normal textures. This bandwidth can be calculated along the dominant frequency direction and its orthogonal directions to accommodate anisotropic textures. The frequency domain energy entropy can be calculated based on the probability distribution after spectral amplitude normalization, using Shannon entropy. Shannon entropy measures the dispersion of spectral energy; low entropy indicates highly concentrated energy (regular textures), while high entropy indicates diffuse energy (random or complex textures).

[0031] S2: The image to be detected is divided into blocks to obtain multiple original image blocks. Based on the local spectral prior map, a personalized frequency domain mask is dynamically generated for each original image block to suppress the frequency band corresponding to the normal texture. The personalized frequency domain mask is applied to the frequency domain representation of the original image block, and a preliminary defect residual map is obtained after amplitude-frequency transformation.

[0032] The image of the spunlace nonwoven fabric to be tested is divided into blocks according to the same block division strategy as the standard image above, resulting in multiple original image blocks.

[0033] For each original image patch, spectral prior information (including dominant frequency direction, energy concentration bandwidth, and frequency domain energy entropy) of its corresponding spatial location is read from the constructed local spectral prior map, and a personalized frequency domain mask is dynamically generated accordingly. This mask accurately suppresses the frequency bands corresponding to normal textures (e.g., regions along the dominant frequency direction and within the energy concentration bandwidth) in the frequency domain, while retaining anomalous frequency domain energy (which may originate from defects or noise). A two-dimensional Fourier transform is performed on the original image patch to convert it to the frequency domain; the personalized frequency domain mask is multiplied point-by-point with its frequency domain representation to suppress normal texture components; then, an amplitude-frequency transform is performed on the spectrum after mask processing (the amplitude is taken and inversely transformed back to the spatial domain, or the amplitude spectrum residual is directly analyzed) to obtain a preliminary defect residual map of the patch, in which normal textures are weakened and potential defect regions are relatively highlighted.

[0034] S3: Perform frequency-amplitude transformation on the preliminary defect residual map to obtain a spectrum map. In response to the fact that the proportion of energy in the mid-low frequency region of the spectrum map to the total energy of the frequency domain exceeds a preset threshold, update the mask parameters and apply the updated mask parameters to the personalized frequency domain mask to obtain the final defect residual map.

[0035] The mask parameters include the frequency band corresponding to the normal texture and the suppression intensity applied to that frequency band to suppress the normal texture components. (See reference...) Figure 2 The mask parameter update method includes steps S30-S32, as detailed below:

[0036] S30: Calculate the local unsuppressed responsivity corresponding to the current suppression intensity to characterize the degree of residual normal texture components in the residual spectrum.

[0037] Local unsuppressed response measures how much normal texture is not effectively suppressed at the current suppression intensity, i.e., the degree of insufficient suppression effect.

[0038] In one embodiment, the method for calculating the local unsuppressed responsivity includes the following steps: under the current mask parameters, obtaining the energy ratio of low frequencies to the full frequency domain in the original image patch as a reference value; applying a perturbation amount to the suppression intensity to generate a perturbation mask; processing the original image patch using the perturbation mask to obtain a residual image; and calculating the ratio of the energy ratio of low frequencies to the full frequency domain in the residual image to the reference value as the local unsuppressed responsivity of the suppression intensity.

[0039] If the local unsuppressed responsivity is close to or equal to 1 (e.g., ≥0.95), it indicates that after increasing the suppression intensity, the proportion of low- and mid-frequency energy in the residual map has hardly decreased. This suggests that the texture components in this region are difficult to suppress further. Possible reasons include: normal textures have been sufficiently suppressed, or the region does not contain regular textures (e.g., defects or noise exist). In this case, the current residual map has already preserved potential anomalous information quite well, and it is not advisable to further increase the suppression intensity.

[0040] If the local unsuppressed responsivity is significantly less than 1 (e.g., ≤0.65), it indicates that enhanced suppression can effectively reduce mid-to-low frequency energy. This region is highly consistent with normal texture characteristics, and the suppression intensity should be further enhanced to improve the background suppression effect.

[0041] Local unsuppressed responsivity is typically less than or equal to 1. If it is greater than 1, it is considered an anomaly or noise interference and is treated as a responsivity of 1. For regions with intermediate values ​​(e.g., 0.65 < local unsuppressed responsivity < 0.95), the adjustment strategy further incorporates the consistency judgment of the neighborhood responsivity: if it differs significantly from the neighborhood, it tends to protect the anomalous signal; if it is consistent with the neighborhood, it can continue to optimize suppression.

[0042] S31: Calculate the optimal adjustment step size based on the local unsuppressed response of the current preliminary defect residual map and the average local unsuppressed response of the preliminary defect residual maps in its neighborhood, and adjust the suppression intensity according to the optimal adjustment step size.

[0043] The method for calculating the optimal adjustment step size includes the following steps:

[0044] The initial step size is set, which refers to the adjustment range of the mask parameters in each iteration. For example, the initial step size can be set in the range of 0.02 to 0.1. If the initial step size is 0.05, it means that the adjustment range of the amplitude is 5% of the total amplitude. The initial step size can be adjusted according to the actual application, such as calibrating it based on factors such as production line speed and texture density.

[0045] Calculate the relative unsuppressed responsivity. In one embodiment, the relative unsuppressed responsivity is calculated by normalizing the absolute difference between the local unsuppressed responsivity and the mean of the neighborhood unsuppressed responsivity. Normalization is a standard technique and will not be elaborated further. The relative unsuppressed responsivity represents the degree to which the normal texture of the current image patch is not sufficiently suppressed, and its deviation from the average level of its neighborhood.

[0046] The neighborhood can be an 8-neighborhood. In practical applications, the range of the neighborhood can be verified and adjusted according to the specific dataset and defect type. In the image edge region, some image patches may not meet the full 8-neighborhood condition. This application dynamically adjusts the neighborhood range in the edge region, using only the actually available neighboring image patches.

[0047] In other embodiments, the relative unsuppressed response can also be calculated by normalizing the ratio of the local unsuppressed response to the mean of the neighboring unsuppressed response as the relative unsuppressed response. Normalization is a standard technique and will not be elaborated further. If the spunlace nonwoven fabric texture on the production line is highly uniform and the defects are mostly localized abrupt changes (such as fiber breaks or holes), the first embodiment (absolute difference) is more suitable; if the product texture exhibits natural fluctuations and the defects are relatively weak or gradual (such as sparse webs or uneven thickness), the second embodiment (ratio) is better at capturing relative anomalies and avoiding missed detections. The two methods are complementary and can be flexibly selected or combined based on actual production line data and defect types.

[0048] If the relative unsuppressed response is greater than a preset response threshold, the product of the initial step size and the relative unsuppressed response is used as the optimal adjustment step size; otherwise, the optimal adjustment step size is 0. The step size is dynamically adjusted based on the relative unsuppressed response to ensure that the adjustment range is neither too large nor too small.

[0049] In one embodiment, the responsivity threshold is set as follows: The standard deviation of the local unsuppressed responsivity of each image patch within the neighborhood of the current image patch is calculated; the responsivity threshold is then set as the product of the standard deviation and a preset multiplier. The standard deviation reflects the degree of fluctuation in the responsivity of each image patch within the neighborhood: a small standard deviation indicates that the neighborhood is generally consistent; a large standard deviation indicates that there are significant differences within the neighborhood itself (there may be defects or texture variations). The standard deviation is multiplied by a preset multiplier (e.g., 2 or 3) to obtain the responsivity threshold.

[0050] Local spectral features are extracted to construct a local spectral prior map. Defects in spunlace nonwovens often manifest as local spectral energy anomalies, but the background texture itself has strong periodicity and random perturbations. Excessive masking suppression leads to the loss of weak defects. Dynamic step size can be finely adjusted in highly sensitive areas to avoid over-smoothing of weak defects.

[0051] S32: Calculate the noise level, adjust the preset initial frequency band according to the noise level to obtain the optimal frequency band, and update the mask parameters according to the optimal frequency band and the suppression intensity.

[0052] In the actual inspection process, the system first sets an initial frequency band for each image block. This frequency band is determined based on the local spectral prior map of the standard defect-free sample, which corresponds to the main energy concentration area of ​​the normal texture of the spunlace nonwoven fabric in the frequency domain.

[0053] The signal-to-noise ratio (SNR) of the current image patch is calculated. The SNR reflects the quality of the image. A higher SNR indicates clearer textures and less noise interference; a lower SNR indicates that the image is more heavily contaminated by noise, and the spectral features may be blurred or distorted.

[0054] When the signal-to-noise ratio (SNR) is high (e.g., greater than 20 dB), the texture is very clear. In this case, the bandwidth is appropriately narrowed to more accurately suppress normal textures and avoid falsely suppressing defects or details. When the SNR is at a medium level (e.g., between 10 dB and 20 dB), the noise effect is considered controllable, and the original initial bandwidth is retained unchanged. When the SNR is low (e.g., below 10 dB), the noise is strong and may cause the texture spectrum to spread or shift. In this case, the bandwidth is appropriately widened to ensure that even under noise interference, the true frequency domain energy range of normal textures can be fully covered to prevent insufficient suppression. The center frequency of the bandwidth remains unchanged and is still determined based on the dominant frequency position in the local spectral prior map; only its bandwidth is adjusted. The resulting new bandwidth is the optimal bandwidth.

[0055] For example, in a standard defect-free sample, the dominant frequency of normal texture in a certain local area is located at the normalized frequency of 0.15. The system's preset initial frequency band is [0.13, 0.17], with a bandwidth of 0.04. This frequency band can cover texture energy well under normal conditions. When appropriate narrowing is needed, the upper and lower boundaries of the frequency band are each shrunk inward by 0.005, resulting in a new optimal frequency band of [0.135, 0.165], with the bandwidth decreasing from 0.04 to 0.03, a narrowing of 25%. For example, assuming that in a standard defect-free spunlace nonwoven fabric sample, the normal texture in a certain local area is mainly concentrated around the normalized frequency of 0.12 in the frequency domain, the system's preset initial frequency band is [0.10, 0.14], that is, a frequency band centered at 0.12 with a total bandwidth of 0.04. This frequency band can effectively suppress normal texture under ideal (low noise) conditions. When a moderate widening is needed, the upper and lower boundaries of the frequency band are each extended outward by 0.01, resulting in a new optimal frequency band of [0.09, 0.15]. The total bandwidth increases from the original 0.04 to 0.06, which is a 50% widening.

[0056] S4: Perform threshold segmentation and / or connected component analysis on the final defect residual map to output the defect location.

[0057] Thresholding converts the final defect residual image (grayscale image) into a binary image (0 represents normal, 1 represents a defect candidate). The principle is to set a segmentation threshold; pixels with residual values ​​higher than this threshold are considered "abnormal" and belong to the defect region. The output is a preliminary binary defect mask, which may contain noise points or fragmented regions. In the thresholding process on the final defect residual image, the threshold can be either an adaptive threshold or a fixed threshold (this is a prior art technique).

[0058] In one embodiment, the entire defect detection process is run on a large number of defect-free samples. The distribution of pixel values ​​in the final defect residual map is statistically analyzed (usually approximating a Gaussian distribution). The threshold is set to the global mean plus three times the standard deviation (i.e., the 3σ principle) to cover most normal fluctuations; any values ​​exceeding this are considered defects. Alternatively, a small number of typical defect samples are labeled, and the minimum response intensity of the defect region in the residual map is observed. A fixed value slightly lower than this intensity is set as a general threshold (e.g., 0.15, with normalized residual values ​​ranging from [0,1]). Fixed threshold calculation is simple and more suitable for scenarios with stable production line lighting, constant fabric tension, and no significant process fluctuations.

[0059] In one embodiment, the adaptive threshold is set dynamically based on the weighted sum of the local standard deviation and the global mean of the final defect residual map. The adaptive threshold is more suitable for scenarios where the production line experiences changes in lighting, fluctuations in fabric tension, uneven texture density, or the presence of weak defects (such as sparse mesh or uneven thickness).

[0060] The weighting method includes: dividing the final defect residual map into blocks to obtain the local standard deviation of each local image block and the global mean of the entire image; adding the local standard deviation and the global mean to obtain the normalized baseline value; using the proportion of the local standard deviation to the normalized baseline value as the weight of the local standard deviation, and using the proportion of the global mean to the normalized baseline value as the weight of the global mean; multiplying the weight of the local standard deviation by the local standard deviation, and multiplying the weight of the global mean by the global mean, and using the sum of the two as the adaptive threshold for the local region.

[0061] Threshold segmentation is performed on the final defect residual map, and optionally, connected component analysis is combined to output the defect location. Alternatively, threshold segmentation and connected component analysis are performed on the final defect residual map to output the defect location. Connected component analysis is an existing technique and will not be described in detail here.

[0062] This application also discloses a defect intelligent identification system for spunlace nonwoven fabric, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the defect intelligent identification method for spunlace nonwoven fabric according to this application is implemented.

[0063] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0064] In this application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (DRAM), dynamic random access memory (DRAM), static random access memory (SRAM), etc., or any other medium that can be used to store desired information and can be accessed by an application program, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.

[0065] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for intelligent identification of defects in a hydroentangled nonwoven fabric, characterized by, The method comprises the steps of: The collected standard non-defective spunlace non-woven fabric image is divided into multiple local image blocks according to spatial positions; a local frequency spectrum prior map corresponding to the spatial positions of the image is constructed; wherein each local image block is aligned in space according to a preset repetition period of the spunlace non-woven fabric surface texture; The to-be-detected image is processed by blocking to obtain multiple original image blocks, and a personalized frequency domain mask is dynamically generated for each original image block based on the local frequency spectrum prior map to suppress the frequency band corresponding to the normal texture; the personalized frequency domain mask is applied to the frequency domain representation of the original image block, and a preliminary defect residual image is obtained through amplitude-frequency transformation; The preliminary defect residual image is subjected to frequency-amplitude transformation to obtain a frequency spectrum image, and in response to the proportion of the medium-low frequency domain energy in the frequency spectrum image exceeding a preset threshold, the mask parameter is updated, the updated mask parameter is applied to the personalized frequency domain mask to obtain a final defect residual image; the mask parameter includes a frequency band corresponding to the normal texture and an inhibition intensity representing the application to the frequency band to suppress the normal texture component; The final defect residual image is subjected to threshold segmentation and / or connected domain analysis to output a defect position; The updating method of the mask parameter is: calculating a local non-suppression response degree corresponding to the current inhibition intensity to represent the residual degree of the normal texture component in the preliminary defect residual image; calculating an optimal adjustment step according to the local non-suppression response degree of the current preliminary defect residual image and the average of the local non-suppression response degrees of the preliminary defect residual image in the neighborhood, adjusting the inhibition intensity according to the optimal adjustment step; calculating the noise degree, adjusting the optimal frequency band according to the preset initial frequency band, and updating the mask parameter according to the optimal frequency band and the inhibition intensity; The calculation method of the local non-suppression response degree comprises the steps of: Under the current mask parameter, the energy proportion of the low frequency and the full frequency domain in the original image block is obtained as a reference value; A disturbance amount is added to the inhibition intensity to generate a disturbance mask, the original image block is processed by using the disturbance mask to obtain a residual image, and the energy proportion of the low frequency and the full frequency domain in the residual image is calculated and compared with the reference value to obtain the local non-suppression response degree of the inhibition intensity; The calculation method of the optimal adjustment step comprises the steps of: setting an initial step; calculating a relative non-suppression response degree; in response to the relative non-suppression response degree being greater than a preset response degree threshold, the product of the initial step and the relative non-suppression response degree is taken as the optimal adjustment step, otherwise the optimal adjustment step is 0; The calculation method of the relative non-suppression response degree is: the absolute difference between the local non-suppression response degree and the average of the neighborhood non-suppression response degrees is normalized as the relative non-suppression response degree; Or the ratio of the local non-suppression response degree to the average of the neighborhood non-suppression response degrees is normalized as the relative non-suppression response degree.

2. The method according to claim 1, wherein The setting method of the response degree threshold is: calculating the standard deviation of the local non-suppression response degrees of each image block in the neighborhood of the current image block; the response degree threshold is set as the product of the standard deviation and a preset multiple coefficient.

3. The method of claim 1, wherein the method further comprises: determining a defect of the spunlace nonwoven fabric based on the image data. The construction of the local frequency spectrum prior map corresponding to the spatial positions of the image comprises: performing two-dimensional Fourier transform on each local image block, extracting local frequency spectrum features to construct the local frequency spectrum prior map.

4. The method according to claim 3, wherein The local spectrum features include a spectrum main frequency direction, an energy concentration bandwidth, and a frequency domain energy entropy.

5. The method of claim 1, wherein the method further comprises: determining a defect of the water-jet nonwoven fabric based on the image. In the threshold segmentation of the final defect residual image, the segmentation threshold is an adaptive threshold or a fixed threshold.

6. The method according to claim 5, wherein The adaptive threshold is set by dynamically setting a weighted sum of a local standard deviation and a global mean value of the final defect residual image.

7. The method of claim 1, wherein the method is characterized by: The noise degree is a signal-to-noise ratio.

8. A defect intelligent identification system for spunlace nonwoven fabric, characterized in that, The method comprises the steps of: The processor and the memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, realize the intelligent identification method of the defect of the spunlace non-woven fabric according to any one of claims 1-7.

Images