Nonwoven fabric defect detection system based on industrial vision

Abstract

The present application relates to the technical field of industrial vision detection, in particular to a non-woven fabric defect detection system based on industrial vision, which comprises a multi-modal imaging device, a background texture suppression device, a defect detection server and a result output terminal; a line scan camera is triggered by interlaced transmission and reflection light sources, and transmission and reflection images are collected under the same physical coordinate system; the background texture suppression device calculates local mutual covariance data, generates a sparse activation mask when a preset trigger condition is met, and intercepts dynamic region of interest data and sends it to the defect detection server; the defect detection server decouples features based on the data, determines the defect state and alarms; through multi-modal interlaced imaging and front-end background texture sparse suppression, the present application effectively reduces the back-end computing power and bandwidth consumption, and improves the defect detection accuracy and response speed under high-speed production lines.

Description

Technical Field

[0001] This invention relates to the field of industrial vision inspection technology, specifically to a nonwoven fabric defect detection system based on industrial vision. Background Technology

[0002] As the application of nonwoven fabric products in medical dressings, hygiene materials, filter materials and industrial wiping continues to expand, the production line has increasingly higher requirements for the consistency of the fabric's surface quality and the efficiency of online inspection. Due to the strong randomness of the fiber arrangement on the surface of nonwoven fabrics, even normal areas will show obvious light and dark fluctuations and texture disturbances. How to accurately identify defects such as holes, foreign objects pressed in, and uneven thickness under high-speed continuous operation has become an important problem in current industrial visual inspection. Traditional nonwoven fabric defect detection currently relies mainly on the following methods: threshold detection based on a single transmission image or a single reflection image, texture analysis based on the entire image, and back-end intelligent recognition directly on the entire image. However, single-image threshold detection, whole-image texture analysis, and full-frame intelligent recognition all have certain drawbacks. For example, single imaging methods are difficult to distinguish between normal random textures and real structural defects, which can easily lead to false alarms or missed alarms. Whole-image texture analysis has a large amount of data and insufficient real-time performance in high-speed production lines. Directly performing backend recognition on the whole image has problems of high bandwidth consumption and heavy computational burden, making it difficult to balance detection accuracy, response speed, and online deployment stability. Summary of the Invention

[0003] The purpose of this invention is to provide a nonwoven fabric defect detection system based on industrial vision, and to solve the following technical problems: Existing nonwoven fabric defect detection technologies have significant shortcomings in accurately identifying defects when faced with strong random texture interference on the material surface and in terms of bandwidth and computing power limitations in high-speed online detection. There is an urgent need to propose a nonwoven fabric defect detection system based on industrial vision that can combine multimodal interleaved imaging with front-end background texture sparsity suppression, dynamically extract abnormal areas for back-end feature decoupling and identification, thereby balancing detection accuracy, response speed and low bandwidth requirements.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows: The system comprises a multimodal imaging device, a background texture suppression device, a defect detection server, and a result output terminal. The multimodal imaging device includes a line scan camera that receives pulse signals from an external production line encoder, as well as a transmitted light source and a reflected light source. The line scan camera is used to interleave the transmitted light source and the reflected light source based on the pulse signals to acquire transmitted image data and reflected image data of the target surface in the same physical coordinate system, and to send the transmitted image data and the reflected image data to the background texture suppression device. The background texture suppression device is used to calculate local cross-covariance data based on the transmission image data and the reflection image data, and when the texture difference represented by the local cross-covariance data meets the preset triggering conditions, generate a sparse activation mask marked with texture abnormal regions, and extract dynamic region of interest data based on the sparse activation mask, and send the dynamic region of interest data to the defect detection server. The defect detection server is used to extract defect features based on the dynamic region of interest data, determine the defect status result of the target object surface, and send a defect alarm command and the defect status result to the result output terminal when the defect status result indicates the existence of a defect; the result output terminal is used to display the defect status result in response to the defect alarm command.

[0005] Furthermore, the background texture suppression device is also used to: extract transmission local gradient data from the transmission image data using a preset sliding window; extract reflection local gradient data from the reflection image data using the preset sliding window; and calculate the local cross-covariance value based on the transmission local gradient data and the reflection local gradient data, as the local cross-covariance data.

[0006] Furthermore, the background texture suppression device is also used to: extract the absolute value of the local cross-covariance data; obtain a preset cross-covariance threshold from the storage space of the background texture suppression device; compare the absolute value with the preset cross-covariance threshold; and determine that the local cross-covariance data satisfies the preset triggering condition in response to the absolute value being greater than the preset cross-covariance threshold. Alternatively, in response to the absolute value being less than or equal to the preset cross-covariance threshold, it is determined that the local cross-covariance data does not meet the preset triggering condition, and the corresponding transmission image data and reflection image data are discarded.

[0007] Furthermore, the background texture suppression device is also used to: when the local covariance data satisfies the preset triggering condition, obtain the original pixel coordinate data of the preset sliding window corresponding to the local covariance data; construct the sparse activation mask based on the original pixel coordinate data; and use the sparse activation mask to truncate the transmission image data and the reflection image data to obtain the dynamic region of interest data.

[0008] Furthermore, the defect detection server is also configured to: input the dynamic region of interest data into a preset sparse perceptron neural network model for feature decoupling; obtain defect prediction probability data output by the sparse perceptron neural network model; determine that the defect status result indicates the existence of a defect in response to the defect prediction probability data being greater than a preset probability threshold; or, determine that the defect status result indicates the absence of a defect in response to the defect prediction probability data being less than or equal to the preset probability threshold.

[0009] Furthermore, the system also includes a model update server; the result output terminal is also used to send the defect status result and the corresponding dynamic region of interest data to the model update server; the model update server is used to update the parameters of the preset sparse perceptual neural network model in the defect detection server based on the defect status result and the dynamic region of interest data.

[0010] Furthermore, the target surface includes a nonwoven fabric surface; the defect status results include at least one of the following: uneven thickness defect data, foreign object indentation defect data, and hole defect data.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. Based on multimodal imaging equipment, the system utilizes the received pulse signals from the external production line encoder to interleave and trigger the transmitted light source and the reflected light source. This design enables the line scan camera to synchronously and separately acquire the transmitted image data and reflected image data of the target surface in the same physical coordinate system. This solution overcomes the problems of missed detection or image misalignment that are easily caused by traditional single light source or asynchronous multi-light source, and ensures strict alignment of multimodal data in physical space. This allows for a more comprehensive coverage and reflection of various defects that may exist on the surface of nonwoven fabric, such as uneven thickness, foreign object indentation, and holes. 2. The background texture suppression device extracts local gradient data from transmission and reflection to calculate local cross-covariance data, thereby accurately quantifying texture differences. When the absolute value of the cross-covariance exceeds a preset threshold, the system generates a sparse activation mask based on the original pixel coordinate data and extracts and sends the data of the dynamically relevant region of interest. Normal region data that does not meet the triggering conditions is discarded directly. This pre-filtering mechanism effectively filters out a large amount of normal non-woven fabric background texture, significantly reduces the transmission of invalid background data, significantly reduces the computing power consumption of subsequent servers, and improves the real-time performance and overall efficiency of industrial vision inspection. 3. The defect detection server utilizes a pre-defined sparse perceptron neural network model to specifically decouple features and predict probabilities for the filtered dynamic region of interest data, effectively improving the accuracy of defect judgment and the efficiency of feature extraction. At the same time, the system introduces a model update server, which can receive the actual defect status results and corresponding dynamic region of interest data returned by the result output terminal, and update the parameters of the sparse perceptron neural network model accordingly. This closed-loop data feedback and model self-learning mechanism enables the system to dynamically adapt to minor changes in non-woven fabric materials or processes in actual production, maintaining long-term, stable, and high-precision defect detection capabilities. Attached Figure Description

[0012] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the workflow of the multimodal defect detection system in the embodiments of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1 A nonwoven fabric defect detection system based on industrial vision includes: a multimodal imaging device, a background texture suppression device, a defect detection server, and a result output terminal; wherein, the multimodal imaging device includes a line scan camera that receives pulse signals from an external production line encoder, as well as a transmitted light source and a reflected light source; the line scan camera is used to trigger the transmitted light source and the reflected light source alternately based on pulse signals, so as to acquire transmitted image data and reflected image data of the target surface in the same physical coordinate system, and send the transmitted image data and reflected image data to the background texture suppression device; The background texture suppression device is used to calculate local cross-covariance data based on transmission and reflection image data. When the texture difference represented by the local cross-covariance data meets the preset triggering conditions, it generates a sparse activation mask marking texture anomalous regions and extracts dynamic regions of interest (ROI) data based on the sparse activation mask, and sends the ROI data to the defect detection server. The defect detection server is used to extract defect features based on the ROI data, determine the defect status of the target object surface, and send a defect alarm command and defect status result to the result output terminal when the defect status result indicates the presence of a defect. The result output terminal is used to display the defect status result in response to the defect alarm command. This embodiment provides a nonwoven fabric defect detection mechanism based on industrial vision. Specifically, the system is deployed at the finished product inspection station of a continuously operating spunlace nonwoven fabric production line to perform online full-width inspection of medical dressing substrates before the fabric roll is wound up. The nonwoven fabric fiber arrangement in the production line is highly random, and normal areas will also appear as a large number of messy brightness and darkness changes in the image. Therefore, the system does not rely on the absolute brightness on a single image to judge defects, but uses the different characterization capabilities of transmission imaging and reflection imaging for the material state at the same location. First, a large area of ​​normal texture is removed at the front end, and then a small number of suspected abnormal areas are sent to the back end for further identification. Specifically, the multimodal imaging device is set above and below the linear detection beam; the line scanning camera is fixedly installed on the imaging base, and the transmitted light source is arranged on the other side of the material to characterize the local basis weight, thickness and porosity of the nonwoven fabric; the reflected light source preferably adopts a low-angle illumination method to highlight surface morphology changes such as surface fiber protrusions, foreign objects pressed in, scratches and fiber roll-ups. Since both types of imaging target the same moving fabric surface, the system receives displacement pulses from the production line encoder and uses displacement rather than time as the sampling reference, so that each row of images is bound to the actual position of the fabric surface. In this way, even if the production line accelerates or decelerates, as long as the encoder still accurately reflects the distance the fabric surface travels, the transmitted image and the reflected image can still be aligned in the same physical coordinate system, avoiding image misalignment caused by speed fluctuations. The line scan camera triggers the transmitted and reflected light sources alternately based on encoder pulses; that is, in consecutive adjacent sampling moments, one sampling corresponds to transmitted illumination, the next sampling corresponds to reflected illumination, and the next sampling switches back to transmitted illumination. The physical significance of this alternation method is that it allows the same photosensitive device to record the response of the same area under two different optical mechanisms sequentially without the need for two independent cameras, thus avoiding registration problems caused by dual-camera installation errors, lens distortion differences, and thermal drift accumulation. For high-speed nonwoven fabrics, this time-division multiplexing sampling based on displacement synchronization is more suitable for continuous online detection. It should be noted that, under staggered triggering, although two adjacent exposures correspond to transmission and reflection modes respectively, their original sampling positions differ by a minimum displacement step along the machine direction. Therefore, in this embodiment, it is not required that the two modes naturally fall on the same row number at the instant of sensor output in the same physical coordinate system. Rather, it means that both are first bound to the same encoder displacement coordinate, and then the system completes rearrangement and registration based on the displacement coordinate. The encoder's minimum displacement step size is set to less than half the length of the target's minimum detectable defect, and is set to a subdivision step size of a single-line spatial resolution. This ensures that the transmission samples corresponding to odd displacement step sizes and the reflection samples corresponding to even displacement step sizes form only a controllable sub-defect level interval in space. Multimodal imaging devices or background texture suppression devices can perform neighbor interpolation, line interpolation, or half-step resampling on the two modal data according to a unified displacement grid, so that the transmission and reflection images are mapped to common coordinate points in the same physical coordinate system in subsequent processing. The system retains the structural simplicity of single-camera interleaved sampling while avoiding analysis bias caused by mistaking adjacent displacement positions as naturally belonging to the same row of pixels. In engineering implementation, a dual-line buffer queue can be used to store the most recent rows of transmission and reflection images. When a reflection row arrives, the system retrieves the adjacent transmission rows from the buffer and reconstructs them according to the corresponding encoder displacement position. For defects such as holes, indentations, thin meshes, and foreign object indentations with length scales significantly larger than the minimum displacement step size, the above reconstruction will not change the determination of whether the defect boundary exists and can significantly improve the stability of the dual-modal correlation analysis. If a product work order requires the detection of finer defects, the encoder pulse subdivision frequency or line scan trigger frequency can be increased simultaneously to keep the two-modal registration error within the allowable range. The background texture suppression device preferably uses an edge computing board with integrated high-speed cache or a field-programmable logic device to receive transmission image data and reflection image data, and perform local correlation analysis on the two. The local covariance data here can be understood in engineering as: on the same small piece of fabric, whether the internal density variation reflected by the transmission channel and the surface undulation variation reflected by the reflection channel show synchronous anomalies or reverse anomalies. Although the random texture of normal fibers is complex, it is often just a messy variation formed by the overlap of ordinary fibers in a small area. Such variations do not necessarily form a stable and significant abnormal correlation between the two modes. Real defects such as holes, hard foreign objects indented, and localized severe thin meshes usually change the light transmittance and surface scattering state of the material at the same time, thus showing a significant abrupt change in local cross-covariance that is different from the normal background. Based on this physical law, the background texture suppression device does not save the entire image for dense calculation. Instead, when it finds that the local cross-covariance meets the preset triggering conditions, it generates a sparse activation mask that only covers the abnormal position, and uses this mask as a basis to extract the dynamic region of interest data from the original image. The dynamic region of interest (ROI) data is sent to the defect detection server. The defect detection server can be an industrial computer, an edge AI server, or an inference node equipped with a graphics processing unit. Its function is not to repeatedly search the entire nonwoven fabric image, but to extract defect features from the dynamic ROI and output the defect status results. The defect status results can include whether the defect exists, the type of defect, the location range of the defect, the severity of the defect, and the corresponding production batch, roll number, and timestamp information. If a defect is determined to exist, the defect detection server sends a defect alarm command to the result output terminal. The result output terminal can be a workshop human-machine interface, a quality control screen, or a workstation audible and visual alarm device to prompt operators to intervene in a timely manner, such as marking the fabric roll, reducing the speed, switching process parameters, or performing slitting and rejection. When encoder pulses are lost, jump, or inconsistent with the trigger frequency of the line scan camera, the system can pause the update of the current coordinate binding relationship and mark the data segment as low-confidence data to avoid continuing to output potentially erroneous defect locations when the displacement reference is inaccurate. When any of the transmitted or reflected light sources exhibits abnormal brightness, flickering instability, or excessive temperature rise, the multimodal imaging device can enter a degraded mode, acquiring only one stable image while simultaneously reporting the incomplete modal state. In this state, the background texture suppression device can stop generating formal alarm results and retain only the original image for manual review to avoid generating a large number of false alarms when optical conditions are out of control. Furthermore, when the background texture suppression device determines that there is no area that meets the trigger conditions during a certain period, it does not send the entire image to the backend but only sends the operating status heartbeat and statistical summary, thereby reducing the transmission load. On a spunlace nonwoven fabric production line used to produce medical wiping materials, the fabric width is over 2 meters, and the production line is running at high speed continuously. At a certain moment, when the fabric surface passes through the field of view of a line scan camera, the transmission image shows that the light transmission in one area suddenly increases, while the reflection image that follows shows that the surface scattering characteristics at the same location have changed significantly. This dual-modal imbalance does not conform to the random fluctuation characteristics of the natural stacking of ordinary fibers, but is closer to the material structure damage caused by holes or foreign object pressure. Therefore, the background texture suppression device only captures this local area, and the defect detection server further confirms that it is an abnormal area and outputs the defect status result. The result output terminal displays the location of the area on the quality inspection interface and prompts the winding station to execute the mark. The purpose of this mechanism is to compress and remove a large number of normal high-entropy textures before they enter the depth recognition process by combining front-end dual-modal imaging with intermediate sparse screening, thereby achieving low latency, low bandwidth usage and high reliability defect detection under high-speed production line conditions.

[0015] The transmission local gradient data is extracted from the transmission image data using a preset sliding window; the reflection local gradient data is extracted from the reflection image data using a preset sliding window; based on the transmission local gradient data and the reflection local gradient data, the local cross-covariance value is calculated and used as the local cross-covariance data. This embodiment provides a mechanism for generating local cross-covariance data. Simply obtaining transmission and reflection images is insufficient to reliably distinguish between normal random textures and real defects. Non-woven fibers are naturally interwoven, and the brightness variation of a single pixel may come from real defects, or from normal fiber thickness variations, feather shadows, or subtle fluctuations in instantaneous illumination. If cross-modal comparisons are performed directly based on pixel brightness, normal textures may be mistakenly identified as anomalies. Therefore, this embodiment further extracts transmission local gradient data and reflection local gradient data within a preset sliding window, and calculates the local cross-covariance value based on the two. Specifically, the preset sliding window can be understood as selecting a small area on the moving fabric surface that progresses with the image, used to observe the spatial trend of grayscale changes within that area, rather than the brightness of a single point; gradient data reflects the direction and intensity of brightness changes from bright to dark and from dark to bright within that area; for transmission images, local gradients mainly map abrupt changes in material density, local thickness boundaries, and the edges of pores; for reflection images, local gradients map more surface bulges, indentations, wrinkles, edges of surface attachments, and fiber fold contours; in other words, the gradient is actually closer to the material structure boundary than the original brightness, and therefore more representative of real physical changes; The background texture suppression device performs joint analysis of transmission and reflection local gradient data within the same sliding window to obtain local covariance values. The covariance value here does not require operators to understand the specific calculation process; its engineering significance lies in determining whether there is a correlation between changes in the internal structure and surface morphology of the same fabric surface that exceeds the normal fluctuation range. Although normal spunlace textures are random, many local variations are merely the natural spread of surface fibers or slight density fluctuations, which may not form synchronous abrupt changes in the two modes. However, real defects often simultaneously disturb both the internal light transmission path and the surface scattering path, thus exhibiting more significant correlation anomalies in the local window. To make the calculation boundaries of this step clearer, both transmission and reflection local gradient data can be obtained from the gray-level difference between adjacent pixels within the window. For any pixel position within the window, gradient components can be extracted along the horizontal and vertical directions respectively, and the gradient magnitude can be used as the local gradient representation of that pixel, for example: , Among them, the coordinates within the window The grayscale value of the transmitted or reflected image at the location is denoted as From this, we obtain This represents the local gradient intensity of transmission or reflection at that coordinate; here... and These correspond to the horizontal and vertical coordinate indices within the window, respectively. and These represent the coordinates of adjacent pixels in the corresponding directions. It represents the absolute value symbol; when a pixel is located at the window boundary and the adjacent pixels in the corresponding direction cannot be obtained, the position can be skipped according to the edge masking rule, and the corresponding difference value can be obtained by using the preset boundary padding method; based on this, the transmission local gradient sequence and the reflection local gradient sequence can be formed respectively. After obtaining the local gradient sequences of the two modalities, the background texture suppression device can calculate their local cross-covariance values ​​within the same window, using the average of the corresponding products after removing the mean as this value, i.e.: , Among them, the first in the same sliding window The transmission gradient value corresponding to each sampling location is denoted as . The corresponding reflection gradient value is denoted as The mean values ​​of the transmission gradient and the mean values ​​of the reflection gradient within this window are denoted as follows: and The number of valid sampling points involved in the calculation is denoted as , For summation, This is the calculated local cross-covariance value. After this processing, if a window only experiences a synchronous increase or decrease in overall brightness without any real structural boundary changes, its cross-covariance value will not be unnecessarily amplified. Conversely, if two modes within a window both exhibit significant abrupt changes at their boundary positions, the value will increase accordingly. It should be noted that the size of the preset sliding window can be set in conjunction with the scale of the defect to be inspected and the resolution of the line scan pixels; if the window is too small, it is prone to oversensitivity to instantaneous changes caused by single fiber hairs; if the window is too large, it may submerge small defects in the surrounding normal texture; therefore, the setting of the window is essentially a balance between detail sensitivity and background robustness to meet the requirements of the minimum inspectable defect size in the production environment; the number of effective sampling points It is preferred to only count the positions of pixels that are not saturated, not underexposed, and not removed by edge masking rules; if the number of valid sampling points in a window is lower than a preset lower limit, the window can be directly marked as unreadable to avoid distortion of cross-covariance results due to insufficient valid data; When there is a significant overall brightness shift on the fabric surface, such as when light pollution causes a slow change in brightness across the entire field of view, the background texture suppression device prioritizes local gradients rather than absolute brightness in the analysis, which can reduce the interference of slowly changing backgrounds on anomaly detection. If the window coverage area happens to be located at the edge of the fabric, a splicing section, or a non-product area, the background texture suppression device can skip the window according to the preset edge shielding rules to avoid the boundary itself being mistakenly identified as a defect. If the transmitted or reflected image is saturated, underexposed, or significantly blurred in a local area, the corresponding window can be directly marked as unreadable, and the subsequent system will decide whether to re-sample or manually review it. On the same medical nonwoven fabric production line, when a roll of material approaches the roll changeover joint, the local fiber spreading state will be more disordered than in the normal stable section. If we still simply judge it by brightness abnormality, we may misjudge normal disturbances near the joint as defects. In this embodiment, the background texture suppression device observes the local gradient relationship of transmission and reflection in each sliding window: if it is just a normal texture caused by random fiber stacking, the two modes usually will not show a stable abnormal correlation. If it is an actual defect such as joint indentation or inclusion of hard particles, the two types of gradients in the window will show structural abrupt changes at the same time, making it easier to be screened out in the subsequent process. The purpose of this step is to transform the high-entropy brightness fluctuations in the original image into more structurally meaningful local boundary information, thereby prioritizing the enhancement of the physical boundaries of real defects and establishing a more reliable front-end basis for subsequent sparse triggering.

[0016] The background texture suppression device is also used for: extracting the absolute value of local cross-covariance data; obtaining a preset cross-covariance threshold from the storage space of the background texture suppression device; comparing the absolute value with the preset cross-covariance threshold; determining that the local cross-covariance data meets a preset trigger condition in response to the absolute value being greater than the preset cross-covariance threshold; and determining that the local cross-covariance data does not meet the preset trigger condition in response to the absolute value being less than or equal to the preset cross-covariance threshold, and discarding the corresponding transmission image data and reflection image data. This embodiment provides a triggering mechanism based on a local cross-covariance threshold. Although the sensitivity to structural boundaries has been improved by using a sliding window and local gradient, a problem still arises in actual production lines: the influence of different types of defects on transmission and reflection is not entirely consistent; some defects will enhance transmission and weaken reflection, such as holes that increase light transmission and reduce surface fiber support; other defects will weaken transmission and enhance reflection, such as foreign objects being pressed in or local fiber agglomeration; therefore, if only the positive and negative directions of local cross-covariance are considered, it is easy to miss defects with opposite physical manifestations but equally important ones. Therefore, this embodiment extracts the absolute value of the local cross-covariance data and compares it with a preset cross-covariance threshold. Specifically, the engineering implication of taking the absolute value is that the system focuses on whether the correlation between the two modes is significant enough, rather than being fixated on whether the change is in the same or opposite direction; as long as the same local area shows abnormal coupling beyond the normal background in both transmission and reflection observation dimensions, it indicates that the area is necessary to enter the subsequent depth recognition; specifically, the tuning process of the preset cross-covariance threshold includes: acquiring transmission and reflection images of a preset number of normal non-woven fabric samples, calculating the local cross-covariance value of each sliding window, and statistically obtaining the mean of the cross-covariance distribution of the normal background texture. with standard deviation and based on Set a preset cross-variance threshold ,in The sensitivity coefficient is set based on the minimum detectable defect specification; When the absolute value of the local cross-covariance data is greater than the preset cross-covariance threshold, the system determines that the location meets the preset trigger condition; this means that the bimodal difference in the area corresponding to the window has exceeded the typical range of normal nonwoven fabric random fluctuations, and has a high probability of being abnormal, so it should be retained and processed in subsequent steps. Conversely, when the absolute value of the local covariance data is less than or equal to the threshold, the background texture suppression device considers the location to be a normal background texture or a weak perturbation that is insufficient to constitute a suspected defect, and discards the corresponding transmission image data and reflection image data; this discarding does not destroy the entire quality traceability chain, but means not using this part as depth recognition input, so as to reduce data handling and inference overhead; if necessary, the system can still retain only volume-level statistical summary or compressed index information; It should be noted that a lower preset cross-variance threshold is not always better. If the threshold is too low, it will introduce a large amount of normal fiber hairs, slight tension fluctuations, and even lighting noise into the back end, weakening the significance of sparse screening. If the threshold is too high, it may miss early minor defects, especially thin webs that have not yet expanded but already pose a quality risk and initial pressure marks. Therefore, this threshold is more suitable as a compromise parameter between production process and detection sensitivity, serving the front-end goal of screening out a small portion of worthwhile areas first. When the detection target is switched to a nonwoven fabric with a different raw material system, such as switching from a regular wiping cloth to a high-loft liquid-absorbing material, the inherent texture noise level of the material will change. If the original threshold is continued to be used at this time, false alarms or missed alarms may occur. To avoid this problem, the background texture suppression device can call different threshold levels according to the product formula or work order number. If there is no corresponding level for the current batch, the system can enter a conservative mode, appropriately increase the retention ratio, and prompt the operator to perform sample calibration. If the cross-covariance result increases abnormally at a certain time period, accompanied by light source pollution alarm or lens contamination alarm, the system can first perform an image quality self-check instead of directly reporting as a defect, so as to avoid mistaking the deterioration of the optical system status as a fabric problem. On this medical nonwoven fabric production line, after a batch of products was switched to a higher bulkiness specification, the local fluctuations in the transmission image itself increased. However, these fluctuations were mostly caused by the natural bulkiness of the fiber layer and did not form a structural anomaly of the same degree in the reflection channel. At this time, the background texture suppression device still used the absolute value of the local cross-covariance as the triggering basis. Only when the bimodal relationship significantly exceeded the threshold corresponding to the product grade was the local area retained. If it was just a normal fluctuation caused by bulkiness, the corresponding window was considered not to meet the triggering condition and was discarded. The purpose of this mechanism is to use the significance of bimodal anomaly correlation as the front-end screening criterion, thereby achieving consistent coverage of defects with different physical manifestations and effectively suppressing the occupation of back-end computing power by random background of nonwoven fabric.

[0017] The background texture suppression device is also used for: obtaining the original pixel coordinate data of the preset sliding window corresponding to the local cross-covariance data when the local cross-covariance data meets the preset triggering conditions; constructing a sparse activation mask based on the original pixel coordinate data; and using the sparse activation mask to extract the transmission image data and the reflection image data to obtain the dynamic region of interest data. This embodiment provides a sparse activation mask construction and dynamic region of interest (ROI) extraction mechanism. Local windows that meet the triggering conditions have already been screened out using threshold criteria. If the entire transmission and reflection images are still transmitted to the backend server, the bandwidth and computing power saved by the front-end screening cannot be truly realized. Especially in high-speed line scanning scenarios, most of the fabric surface is still a normal area, and continuing to retain the entire image will result in redundant data transfer. Therefore, this embodiment further constructs a sparse activation mask based on the original pixel coordinate data of the sliding window that meets the triggering conditions, and extracts the dynamic ROI data accordingly. Specifically, the original pixel coordinate data essentially corresponds to the digital mapping of the actual fabric position in the production line. Since both the transmission and reflection images are acquired in a unified physical coordinate system, once a sliding window is identified as abnormal, its position in the image can be stably mapped to the specific horizontal and vertical segments of the fabric. Based on these coordinates, the background texture suppression device constructs a sparse activation mask. The sparse activation mask can be understood as an index map that only marks the positions worth continuing to look at, while the remaining large normal areas remain inactive. The system does not need to use the entire image as input to the backend neural network, but only extracts necessary local image patches around these activation positions. When using sparse activation masks to crop transmission and reflection image data, the dynamic region of interest is not limited to a fixed size; different defects have different physical scales and expansion directions on the fabric surface; thin mesh and pinhole defects are often relatively localized, while scratches, strip indentations, or continuous cloud spots may extend along the machine direction; the dynamic aspect in this embodiment is manifested in that: the cropped area can be merged, expanded, or clipped according to the spatial distribution of the activation area, the connectivity between adjacent activation windows, and the defect context range of concern to the process, so that the back end can see both the defect body and its surrounding background, which is beneficial for distinguishing between real defects and normal texture boundaries; It should be noted that, under the single-line scanning camera interleaved sampling structure corresponding to this embodiment, the original pixel coordinate data is preferably not simply using the original row number of the sensor, but using a combination of horizontal pixel index and encoder displacement index; for the case where the transmission mode and the reflection mode fall on the odd and even displacement indices respectively, the background texture suppression device first maps the sliding window that meets the triggering condition to a unified displacement coordinate axis, and then generates a sparse activation mask on the unified coordinate axis; The mask marks the physical location of the fabric, not the original row number position unique to a certain mode. When capturing the dynamic region of interest later, the system can retrieve the corresponding image block from the transmission buffer and the reflection buffer respectively based on this unified coordinate. If there is no original sampling row that is completely in the same position at a certain coordinate, the adjacent displacement row is called for padding or interpolation reconstruction. The sparse activation mask can continue to play the role of transmitting anomalies, and will not cause the two modal blocks to be misaligned by half a beat in the machine direction due to interleaved sampling. To avoid the backend receiving semantically incomplete data at the same abnormal location, the dynamic region of interest (ROI) data preferably includes at least: a transmission image patch, a reflection image patch corresponding to the same unified coordinate range, and coordinate index information bound to the patch; when more context needs to be retained, it can also include the corresponding sparse activation mask local fragment, truncation boundary information, and trigger window identifier; with this definition, the dynamic ROI data received by the backend server is not an isolated patch of a single modality, but a group of data organized around the same abnormal physical region, thus maintaining consistency with the sampling and filtering logic under the same physical coordinate system mentioned above; In engineering implementation, sparse activation masks can also serve as an intermediate carrier for subsequent quality traceability. Because each activation area naturally carries image coordinates, roll number, encoder displacement position and timing information, when the back-end server outputs the defect status results, it can directly trace back its position in the whole roll of material for rejection, switching process parameters or establishing a defect distribution map. When multiple windows that meet the triggering conditions are too densely packed in space, the background texture suppression device can first merge adjacent windows to avoid repeatedly capturing multiple highly overlapping small blocks of data for the same defect; if the activated area happens to fall on the edge of the fabric, the curling edge fluctuation area, or the known process clamping area, the system can trim this part according to the preset area shielding rules to prevent non-product boundaries from being brought into the backend identification; if the number of triggering areas increases abnormally and exceeds the real-time processing capacity of the backend server, the background texture suppression device can first retain high-confidence areas according to the priority strategy, and cache and queue low-priority areas or only record the index to prevent instantaneous congestion from causing frame drops; In a roll of spunlace nonwoven fabric produced on this line, a section of indentation extending along the running direction appears on the fabric surface, accompanied by several scattered small bright spots nearby. If each window is captured independently, not only will a large amount of duplicate data be generated, but it will also be difficult for the backend to determine whether these anomalies belong to the same defect event. In this embodiment, the background texture suppression device merges the spatially continuous active windows into a longer dynamic region of interest, while keeping isolated bright spots far from the main indentation captured separately. In this way, the backend can see the continuous direction of the indentation and classify the scattered anomalies separately. The purpose of this step is to truly transform the front-end judgment of anomaly detection into a data organization method that only transmits anomalies, thereby achieving a data sparse closed loop between imaging, screening and recognition, and reducing the ineffective burden on the back-end model. Example

[0018] The defect detection server is also used for: inputting dynamic region of interest data into a preset sparse perceptron neural network model for feature decoupling; obtaining defect prediction probability data output by the sparse perceptron neural network model; determining the defect status result indicating the presence of a defect in response to the defect prediction probability data being greater than a preset probability threshold; and determining the defect status result indicating the absence of a defect in response to the defect prediction probability data being less than or equal to the preset probability threshold. This embodiment provides a defect determination mechanism based on a sparse perceptual neural network model. The front end has significantly reduced the range of the region to be analyzed through sparse activation masks, but the dynamic region of interest sent to the back end may still contain many cases that are not defects but are not defects, such as normal fiber agglomeration, local gloss changes, joint edge disturbances, or slight thickness fluctuations within the allowable range of the process. If the front end triggers an alarm directly based on the threshold, it is easy to cause false alarms. Therefore, this embodiment introduces a preset sparse perceptual neural network model to decouple the dynamic region of interest data and complete the final determination based on the output defect prediction probability data. Specifically, feature decoupling in this embodiment involves the model identifying components belonging to random background texture, components belonging to imaging condition perturbations, and components that truly reflect material defects in the region. For nonwoven fabrics, many normal areas may appear visually cluttered, but this clutter is acceptable and normal. Real defects, on the other hand, typically exhibit disruptions to material continuity, density distribution, or surface topology. Since the sparse perceptual neural network model only processes a small number of dynamically relevant regions of interest that have been screened out by the front end, it can retain more contextual information with limited computing power to understand the relationship between defect boundaries and surrounding textures, instead of distributing computing power evenly to a large area of ​​normal background as in traditional whole-image detection. To make the term sparse sensing clearer in engineering, sparse sensing neural network models preferably receive groups of dynamic regions of interest as input, rather than receiving only a single patch; the input includes at least transmission image patches and reflection image patches corresponding to the same anomalous region; When enhanced positioning stability is required, sparse activation mask local segments, coordinate index information of the cropped region, and modal identifiers representing the source of the crop can also be attached. The model encodes transmission and reflection information in the front end, retaining their respective responses to changes in the internal structure and surface morphology of the material. Specifically, the front-end encoder of the sparse perceptual neural network model includes parallel two-stream convolutional layers to extract shallow texture features of transmission image crop and reflection image crop, respectively. In the intermediate fusion layer, the spatial correlation weights of the transmission feature map and the reflection feature map are calculated through a cross-modal attention mechanism, and the dual-stream features are multiplied by matrix and weighted concatenated to achieve the fusion of dual-modal features. Then, the two types of information are fused in the intermediate layer, and the slow brightness drift, random texture of ordinary fibers and local imaging perturbations that are not related to defects are separated as much as possible. This feature decoupling process distinguishes the feature components that are related to defect determination and those that are not related to defect determination in terms of spatial dimension and channel dimension within the network, extracts the target features for defect determination, and then uses the part of the features that are more directly related to defects to participate in the subsequent determination. For example, the sparse perceptual neural network model can use the ResNet series as the basic backbone network, and the cross-entropy loss function is used to optimize the network parameters iteratively during the training phase. After the model outputs defect prediction probability data, the defect detection server compares it with a preset probability threshold. When the defect prediction probability data is greater than the threshold, it indicates that the dynamic region of interest is closer to the statistical and physical characteristics of a real defect in the model's view, thus determining that the defect status indicates the presence of a defect. Conversely, when the defect prediction probability data is less than or equal to the threshold, it is considered that the region is more likely to belong to normal texture fluctuations, permissible process disturbances, or imaging noise, thus determining that the defect status indicates the absence of a defect. The probability threshold here does not mean a simple empirical score, but rather serves a quality management strategy: for high-risk products such as medical materials, the threshold can be set more conservatively to reduce missed detections; for ordinary industrial wiping materials, different balance points can be used between missed detections and false alarms. The defect prediction probability data can be a probability value directly output by the model regarding the existence of a defect; or it can be the total probability of a defect obtained by further summing the scores corresponding to each defect category when the model outputs multiple defect results simultaneously. In the latter case, the defect detection server preferably converts the model output into a single defect prediction probability data according to a unified rule before comparing it with a preset probability threshold, thereby avoiding inconsistencies in judgment criteria due to different output formats under different product work orders. Correspondingly, in addition to the presence or absence of defects, the defect status result can also include the defect category and level to guide different handling methods. It should be noted that the input of the sparse perceptual neural network model can include transmission image patches, reflection image patches, and necessary location information or contextual labels, thereby preserving the discriminative advantages brought by the bimodality; in addition to the presence or absence of defects, the defect status result output by the model can also be further accompanied by defect category and level to guide different handling methods. When the data quality of the dynamically generated region of interest is insufficient, such as when the image is severely blurred, locally overexposed, the cropped area is too small, or it crosses the edge of the image, the defect detection server may not directly give a clear conclusion of whether the image is defective or not. Instead, it may output a status pending review and push the corresponding area to the manual review queue. If the model output probability is close to the threshold, which is a case of unclear boundaries, the system can also adopt a dual strategy: on the one hand, do not trigger strong alarms for the time being, and only generate soft prompts; on the other hand, archive the region along with its context for subsequent model updates; if the backend server is temporarily overloaded, dynamic regions of interest with higher confidence assigned by the frontend can be processed first, and low-priority regions are deferred in the queue to ensure that core anomalies are not delayed. When the system is in the aforementioned modal incomplete degradation state, the defect detection server prefers not to directly use the single-modal dynamic region of interest as the basis for formal defect rejection, but instead outputs a low-confidence conclusion or a state pending review associated with the degradation state; the front-end strategy of stopping formal alarms when optical conditions are out of control can be consistent with the back-end output strategy of probability judgment, so as to avoid contradictory processing results of the same anomaly at different levels of the system. In this medical nonwoven fabric production line, a certain dynamic region of interest appears as a slight dark band in the transmission image and a thin, bright line in the reflection image. Based on process experience, this could be a longitudinal scratch or simply a residual texture from hydroentanglement. The front end has already screened it out, but it has not triggered an alarm directly. After the defect detection server inputs this region into the sparse perceptual neural network model, the model makes a judgment based on comprehensive features such as whether the bright line is consistent with the direction of the surrounding fibers, whether the dark band continuously crosses multiple local contexts, and whether the reflection anomaly co-occurs with the transmission anomaly. Only when the defect prediction probability exceeds a set threshold is it output as a real scratch. The purpose of this mechanism is to further refine the distinction between suspected anomalies and actual defects based on the sparse screening at the front end, thereby achieving online quality assessment with low false alarms, categorizability, and traceability.

[0019] The system also includes a model update server; the result output terminal is also used to send defect status results and corresponding dynamic region of interest data to the model update server; the model update server is used to update the parameters of the preset sparse perceptual neural network model in the defect detection server based on the defect status results and dynamic region of interest data. This embodiment provides a closed-loop update mechanism for the model. The sparse perceptual neural network model is already able to determine defects in dynamic regions of interest. However, nonwoven fabric production is not a static environment. Changes in the length and ratio of raw material fibers, bonding methods, hydroentanglement strength, drying conditions, and equipment wear can all cause the slow migration of normal texture distribution and defect appearance. If the model is not updated for a long time, even if the initial training effect is good, it may become insensitive to new defects and have more false alarms for new background textures. Therefore, this embodiment further sets up a model update server, and the result output terminal sends back the defect status results and their corresponding dynamic regions of interest data to update the parameters of the preset sparse perceptual neural network model in the defect detection server. Specifically, in addition to real-time display and alarm functions, the result output terminal can also serve as an interactive entry point for manual review and quality confirmation. After viewing the defect status results provided by the system, operators can combine physical sampling inspection, offline laboratory review, roll-to-roll splitting results, or client feedback to confirm, correct, or supplement annotations for certain areas. After receiving these dynamic regions of interest data with actual quality conclusions, the model update server can incorporate them into the incremental training sample library. In this way, the model learns not only the existing defect patterns in the initial dataset, but also continuously absorbs new defects, new backgrounds, and new disturbances that actually occur on the current production line. To maintain technical consistency between the online judgment link and the model update link, the defect status results sent by the result output terminal to the model update server preferably include the system's automatic judgment results and the manually confirmed status associated with those results. In other words, the model update server does not use all online prediction results indiscriminately for parameter updates, but instead prioritizes using dynamic region of interest data that has been manually reviewed, physically inspected, or verified by downstream quality conclusions as valid update samples. For samples that have only been initially judged by the machine but have not yet been confirmed, they can be temporarily stored in the candidate sample area, and it will be decided whether to include them in the training after the labels are completed. This does not change the overall link of parameter updating based on defect state results and dynamic region of interest data, and avoids incorrect labels directly polluting the backend model. Furthermore, the model update server can adopt offline or quasi-online update methods. In the offline update method, the system collects a certain number of confirmed samples by shift, batch, or work order and updates the model uniformly. After verification, the model is sent to the defect detection server. In the quasi-online update method, for new samples with high reference value and strong representativeness, the model update server can prioritize short-cycle fine-tuning to adapt to raw material fluctuations or equipment status changes more quickly. Regardless of the update method used, the updated model should preferably be tested with a verification sample corresponding to the current production product to confirm that the false alarm rate and false negative rate meet the process requirements before officially replacing the online model. It is important to emphasize that the model update in this embodiment does not change the physical triggering mechanism of the front end, but continuously corrects how the back end understands the screened abnormal regions; this layered update method can maintain the real-time performance of sparse triggering of the front end, while gradually improving the classification and recognition capabilities of the back end. In engineering deployment, the model update server preferably manages the update process using versioning; that is, each parameter update is bound to the corresponding sample batch, product work order, training time, and validation results; when the defect detection server receives a new model, it can first conduct a trial run in a controlled window using parallel comparison or gray-scale replacement. After confirming that the difference in judgment between the old and new models under the same product conditions is within the allowable range, it can then be switched to the official online version. This makes the linkage between the model update server, the defect detection server, and the result output terminal clearer and avoids mutual interference between the model replacement process and the online judgment process. When the manually corrected data received by the output terminal is contradictory, or the sample labeling source is unreliable, the model update server can isolate this part of the data as samples to be reviewed and not use it directly for parameter updates, so as to avoid the model being contaminated by erroneous labels. If the performance of a model decreases in the validation set after a certain model update, or the false alarms increase abnormally during online trial operation, the system can automatically roll back to the previous stable version. If a network communication failure causes the model update server to be unable to receive the data returned by the output terminal in a timely manner, the defect detection server will maintain the current model and continue to run, and cache the new samples locally until communication is restored before synchronizing. After the medical nonwoven fabric production line had been running stably for several weeks, the upstream supplier switched some fiber raw materials, and the reflective texture of the normal products was slightly rougher than before. This caused the backend model to generate more soft alarms for certain normal areas. The quality inspectors checked these areas on the result output terminal and marked them as normal textures. The model update server included these new samples along with real holes and foreign objects indented during the same period in the update. After verification, the updated model was redeployed to the defect detection server, making the system more adaptable to the new raw material background while still remaining sensitive to the identification of real defects. The purpose of this mechanism is to build a closed-loop learning chain from online detection and manual verification to model iteration, thereby achieving adaptive optimization in the long-term operation of the system and reducing performance degradation caused by process and raw material drift.

[0020] The target surface includes the non-woven fabric surface; the defect status results include at least one of the following: uneven thickness defect data, foreign object indentation defect data, and hole defect data. This embodiment provides a defect type landing mechanism for nonwoven fabric surfaces; specifically, the target object in the aforementioned embodiments is preferably a continuously running nonwoven fabric web, which is especially suitable for products that require appearance integrity and structural uniformity, such as medical dressing substrates, sanitary material base fabrics, filter material base fabrics, and industrial wiping cloths; the defect status results may include at least one of the following: uneven thickness defect data, foreign object indentation defect data, and hole defect data, and may also output multiple types and their level information simultaneously; Specifically, uneven thickness defects are usually physically manifested as abnormal local weight distribution, excessively thick fiber layers, or excessively thin local layers. In transmission images, they are often more easily reflected as differences in light transmission, while in reflection images, they may be accompanied by changes in the uniformity of surface scattering. For this type of defect, the system is not only concerned with single points that are too bright or too dark, but also with the disruption of the structural continuity within a certain area. This is because the normal random texture of nonwoven fabrics naturally has light and dark undulations. Only when these undulations reach a level that affects the product's absorbency, strength, or uniformity does it constitute uneven thickness in a quality sense. Foreign object indentation defects correspond to foreign hard particles, metal fragments, resin lumps, or process residues embedded in the fabric surface. The typical characteristic of this type of defect is that it simultaneously alters the local light transmission path and surface undulation: in the transmitted image, abnormal occlusion or local high-density shadows may appear, while in the reflected image, it appears as a raised or indented area with sharp edges and obvious light-dark contrast. This type of defect is particularly sensitive for medical and hygiene products because it not only affects the appearance but may also pose a safety risk in use. Hole defects correspond to areas where the continuity of the material is truly interrupted, including pinholes, tear initiation points, small holes, and areas where fibers are completely missing. Its most significant physical feature is an abnormally enhanced light-transmitting area in the transmission image, while the reflected image often shows changes in edge morphology, fiber breakage boundaries, or scattering characteristics at the same location. Since holes directly affect the material's barrier properties, mechanical strength, and finished product yield, they are generally classified as priority alarms. Furthermore, in addition to the type, the defect status results can also include the defect's lateral location, longitudinal length, area level, displacement information bound to the roll number, and alarm level; for example, continuous large-area thickness unevenness can be marked as a process fluctuation risk, prompting process personnel to check the web forming and hydroentangling parameters; foreign matter intrusion can be marked as equipment hygiene or upstream material issues, prompting investigation of the source; holes can be marked as serious defects, directly triggering roll segment rejection or shutdown for inspection; When a dynamic region of interest simultaneously exhibits multiple defect characteristics, such as a foreign object pressing into the surrounding area causing a local hole, the defect detection server can output a composite defect status result instead of forcibly classifying it into a single category. If the current evidence is insufficient to reliably distinguish between uneven thickness and slight process fluctuations, the system can temporarily record it as a structural anomaly pending verification to avoid misclassification affecting process judgment. If a certain type of defect has never appeared in the current batch and the model confidence is significantly insufficient, the result output terminal can only indicate the location of the anomaly without directly giving the specific category, allowing quality inspectors to confirm it in conjunction with the actual product. In the aforementioned medical spunlace nonwoven fabric production scenario, a section of fabric is first triggered by the front end and sent to the back end for identification. If the model, combining transmission and reflection local features, determines that the area only exhibits local fiber density abnormalities and the range is relatively gradual, then the thickness unevenness defect data is output. If the same type of area simultaneously presents a sharp indentation outline in the reflection image, then it is more likely to output foreign object indentation defect data. If the transmission channel shows obvious light transmission enhancement and the edge is accompanied by a fracture morphology, then the hole defect data is output and a high-priority alarm is triggered. The purpose of this mechanism is to ultimately translate the aforementioned imaging, screening, and identification process into defect semantics that can be directly used in the nonwoven fabric industry, thereby achieving an executable conversion from image anomalies to process quality conclusions.

[0021] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A nonwoven fabric defect detection system based on industrial vision, characterized in that, include: The system comprises a multimodal imaging device, a background texture suppression device, a defect detection server, and a result output terminal. The multimodal imaging device includes a line scan camera that receives pulse signals from an external production line encoder, as well as a transmitted light source and a reflected light source. The line scan camera is used to interleave the transmitted light source and the reflected light source based on the pulse signals to acquire transmitted image data and reflected image data of the target surface in the same physical coordinate system, and to send the transmitted image data and the reflected image data to the background texture suppression device. The background texture suppression device is used to calculate local cross-covariance data based on the transmission image data and the reflection image data, and when the texture difference represented by the local cross-covariance data meets the preset triggering conditions, generate a sparse activation mask marked with texture abnormal regions, and extract dynamic region of interest data based on the sparse activation mask, and send the dynamic region of interest data to the defect detection server. The defect detection server is used to extract defect features based on the dynamic region of interest data, determine the defect status result of the target object surface, and send a defect alarm command and the defect status result to the result output terminal when the defect status result indicates the existence of a defect; the result output terminal is used to display the defect status result in response to the defect alarm command.

2. The nonwoven fabric defect detection system based on industrial vision according to claim 1, characterized in that, The background texture suppression device is further configured to: extract transmission local gradient data from the transmission image data using a preset sliding window; extract reflection local gradient data from the reflection image data using the preset sliding window; and calculate the local cross-covariance value based on the transmission local gradient data and the reflection local gradient data, using it as the local cross-covariance data.

3. The nonwoven fabric defect detection system based on industrial vision according to claim 2, characterized in that, The background texture suppression device is further configured to: extract the absolute value of the local cross-covariance data; obtain a preset cross-covariance threshold from the storage space of the background texture suppression device; compare the absolute value with the preset cross-covariance threshold; and determine that the local cross-covariance data satisfies the preset triggering condition in response to the absolute value being greater than the preset cross-covariance threshold. Alternatively, in response to the absolute value being less than or equal to the preset cross-covariance threshold, it is determined that the local cross-covariance data does not meet the preset triggering condition, and the corresponding transmission image data and reflection image data are discarded.

4. The nonwoven fabric defect detection system based on industrial vision according to claim 2, characterized in that, The background texture suppression device is further configured to: when the local covariance data satisfies the preset triggering condition, obtain the original pixel coordinate data of the preset sliding window corresponding to the local covariance data; and construct the sparse activation mask based on the original pixel coordinate data. The sparse activation mask is used to extract the transmission image data and the reflection image data to obtain the dynamic region of interest data.

5. The nonwoven fabric defect detection system based on industrial vision according to claim 1, characterized in that, The defect detection server is further configured to: input the dynamic region of interest data into a preset sparse perceptron neural network model for feature decoupling; obtain defect prediction probability data output by the sparse perceptron neural network model; determine that the defect status result indicates the existence of a defect in response to the defect prediction probability data being greater than a preset probability threshold; or, determine that the defect status result indicates the absence of a defect in response to the defect prediction probability data being less than or equal to the preset probability threshold.

6. The nonwoven fabric defect detection system based on industrial vision according to claim 5, characterized in that, The system also includes a model update server; the result output terminal is further used to send the defect status result and the corresponding dynamic region of interest data to the model update server; the model update server is used to update the parameters of the preset sparse perceptual neural network model in the defect detection server based on the defect status result and the dynamic region of interest data.

7. The nonwoven fabric defect detection system based on industrial vision according to any one of claims 1 to 6, characterized in that, The target surface includes a nonwoven fabric surface; the defect status results include at least one of the following: uneven thickness defect data, foreign object indentation defect data, and hole defect data.

Images