Composite film quality evaluation system based on machine vision

Abstract

The invention discloses a composite film quality evaluation system based on machine vision, and the system comprises an image collection module which is used for synchronously collecting a transmission image and a reflection image, and completing the correction, normalization and registration processing; the feature extraction module is used for performing convolution feature extraction on the image to generate transmission features and reflection features; the defect positioning module is used for inputting the transmission features and the reflection features into the bimodal defect positioning module based on LPC-Net for feature fusion; the confidence generation module is used for jointly calculating the thickness estimation graph and the defect initial segmentation result to generate a pixel confidence graph; the defect segmentation and time sequence module is used for executing motion compensation and time sequence correction on continuous frames to obtain a time sequence correction result; the feature statistics module is used for calculating a thickness uniformity index; and the quality evaluation module is used for generating a composite film quality score and a quality grade and sending a result to production line control equipment. According to the invention, quality evaluation of the composite film is realized.

Description

Technical Field

[0001] This invention relates to the field of industrial online inspection technology, and in particular to a composite film quality assessment system based on machine vision. Background Technology

[0002] Composite films, as important substrates in flexible packaging, battery separators, and optical functional films, are typically produced using continuous processes such as extrusion, stretching, and lamination. Quality issues such as surface defects, internal air bubbles, and uneven thickness directly affect the product's mechanical properties, barrier properties, and appearance. Current production lines mostly rely on manual visual inspection or single optical inspection equipment for sampling quality control of composite films. Manual inspection suffers from high subjectivity, low efficiency, and difficulty in achieving full coverage of continuous large rolls of products. While some lines have introduced industrial cameras and photoelectric sensors, these are often limited to single-sided reflection imaging or simple light transmission detection, resulting in limited information dimensions and making it difficult to identify both surface defects and internal structural anomalies.

[0003] In the field of machine vision inspection, existing technologies often employ traditional image processing methods based on threshold segmentation, edge detection, and connected component analysis, or use simple convolutional neural networks to classify defects in local image blocks. In composite thin film scenarios with high background ratio, small defect size, and complex texture, these solutions often suffer from the problem of missing small scratches, bubbles, pinholes, and falsely detecting noise points.

[0004] In addition, most existing algorithms process each frame of image independently without taking into account the temporal characteristics of the continuous movement of the composite film along the production line, and without constraining the temporal consistency of the segmentation results between adjacent frames. This makes it easy to produce "flickering" defect results when there are fluctuations in lighting, noise interference, or local texture changes, which affects the stable tracking of long strip defects and periodic defects.

[0005] Therefore, how to provide a machine vision-based composite film quality assessment system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a machine vision-based composite film quality assessment system. This invention improves the detection accuracy of small and internal defects by combining transmission and reflection dual-modal imaging with two-stage segmentation based on LPC-Net. At the same time, it introduces thickness estimation and pixel confidence joint modeling and temporal consistency constraints to reduce false detections and missed detections and enhance the stability of results. Based on this, a quality scoring and grade evaluation mechanism is constructed to realize quantitative evaluation of composite film quality and real-time closed-loop adjustment of production line processes.

[0007] A machine vision-based composite film quality assessment system according to an embodiment of the present invention includes: The image acquisition module is used to simultaneously acquire the transmission and reflection images of the composite thin film and to complete the correction, normalization and registration processing. The feature extraction module is used to perform convolutional feature extraction on the preprocessed transmission and reflection images to generate transmission and reflection features. The defect localization module is used to input transmission and reflection features into the LPC-Net-based dual-modal defect localization module, perform feature fusion, and output defect candidate regions and initial defect segmentation results. The confidence generation module is used to obtain a thickness estimation map based on the transmittance calibration, and to jointly calculate and generate a pixel confidence map by combining the thickness estimation map with the initial defect segmentation results. The defect segmentation and temporal module is used to fuse defect candidate regions and pixel confidence maps to generate defect segmentation results, and to perform motion compensation and temporal correction on consecutive frames to obtain temporal correction results. The feature statistics module is used to count the number, area and distribution characteristics of defects based on the time-series correction results, and to calculate the thickness uniformity index based on the thickness estimation map. The quality assessment module is used to construct a quality assessment feature vector, generate a composite film quality score and quality grade, and send the results to the production line control equipment.

[0008] Optionally, modules can be integrated using the following methods: The transmission and reflection images of the composite thin film were acquired, and feature extraction was performed to obtain the transmission features and reflection features, respectively. Transmission and reflection features are input into the LPC-Net-based dual-modal defect localization module to perform cross-modal feature fusion, and defect candidate regions are determined based on the defect localization feature map to obtain the initial defect segmentation result. Transmittance calibration is performed on the transmission image to generate a thickness estimation map. The thickness estimation map and the initial defect segmentation result are simultaneously input into the structure-thickness joint pixel confidence module based on LPC-Net to generate a pixel confidence map. The defect candidate region is fused with the pixel confidence map to generate the defect segmentation result of the composite film; Temporal consistency constraints are applied to the continuous sampling frames of the composite film. The defect segmentation results of adjacent frames are matched with motion compensation according to the direction of production line movement. The defect segmentation results are corrected based on the pixel response differences of adjacent frames to obtain the temporal correction results. The defect distribution characteristics of the composite film are determined based on the time-series correction results, and the thickness uniformity index of the composite film is determined based on the thickness estimation diagram. The quality assessment feature vector is constructed based on the thickness uniformity index, and a composite film quality score is generated according to the preset weighting rules to determine the quality level of the composite film. The score is then sent to the production line control equipment for process adjustment and quality early warning.

[0009] Optionally, obtaining the transmission and reflection characteristics specifically includes: On both sides of the inspection area of ​​the composite film production line, a transmission imaging light source and a transmission industrial camera are respectively set up, and a reflection imaging light source and a reflection industrial camera. The transmission imaging light source and the transmission industrial camera are arranged opposite each other to collect the transmission image of the composite film. The reflection imaging light source and the reflection industrial camera are arranged with preset incident angle and observation angle to collect the reflection image of the composite film. The camera exposure time and trigger frequency are set according to the running speed of the composite film so that the transmission image and the reflection image are collected synchronously at the same physical position. Simultaneous acquisition of transmitted and reflected images, along with image preprocessing, including geometric distortion correction based on camera calibration parameters, normalization of overall brightness and contrast to reduce uneven illumination and noise, and spatial registration of transmitted and reflected images to ensure that the two images correspond to the same composite film region at pixel positions, resulting in preprocessed transmitted and reflected images. Feature extraction is performed on the preprocessed transmission and reflection images. Convolution, nonlinear activation and downsampling are performed sequentially to extract transmission and reflection features that characterize the internal structure and thickness variation of the composite film.

[0010] Optionally, obtaining the initial defect segmentation result specifically includes: Transmission and reflection features are input into a dual-modal defect localization module based on LPC-Net. Within the LPC-Net-based dual-modal defect localization module, different feature vectors at the same spatial location are weighted and summed according to preset weight parameters, and channel stitching is performed. Convolution and nonlinear activation operations are then executed sequentially to obtain a fused feature map that characterizes the defect response of the composite film. The LPC-Net is a two-stage industrial defect segmentation deep learning model consisting of a defect localization stage and a pixel confidence estimation stage, which is used to achieve high-precision defect segmentation in images with high background ratio and small defect area size. The fused feature map is input into the defect localization sub-network based on LPC-Net. Multi-layer convolution operation, nonlinear live operation and downsampling operation are performed in the defect localization sub-network in sequence to extract features and spatially compress the fused feature map, generating a defect localization feature map to characterize the defect response intensity at each position of the composite film. Based on the defect localization feature map, the initial segmentation probability of the current position belonging to the defect category is calculated for each spatial position to obtain the corresponding initial segmentation probability distribution. The initial segmentation probability distribution is subjected to threshold segmentation processing. A preset probability threshold is used to distinguish between defect pixels and background pixels. Positions with an initial segmentation probability greater than or equal to the probability threshold are marked as defect candidate pixels, and otherwise marked as background pixels. Defect candidate regions are generated based on the spatial connectivity of the defect candidate pixels. The defect candidate regions and the initial segmentation probability distribution are used together as the initial defect segmentation result for output.

[0011] Optionally, the generation of the pixel confidence map specifically includes: Transmittance calibration is performed on the preprocessed transmission image. Different composite film calibration samples with known thicknesses are selected, and different calibration regions are divided. The actual thickness value and the average gray value in the transmission image are recorded for each calibration region. Based on the correspondence between the actual thickness value and the average gray value, a calibration relationship between transmission gray value and local thickness of composite film is established. The preprocessed transmission image of the composite film under test is input into the calibration relationship pixel by pixel. According to the calibration relationship, the gray value of each pixel is converted into the corresponding thickness estimate value. Thickness estimation is performed for all pixels of the entire image to generate a thickness estimation map that represents the local thickness distribution of composite film in space. The initial defect segmentation result and the thickness estimation map are input into the structure-thickness joint pixel confidence module based on LPC-Net. The structure-thickness joint pixel confidence module belongs to the pixel confidence sub-network in the two-stage network of LPC-Net. By combining the initial segmentation probability distribution with the thickness estimation value and local thickness change information at the pixel level, a joint feature vector including defect response, thickness estimation and neighborhood context is constructed for each position. The confidence level of the defect category to which each pixel belongs is re-estimated, forming a refined confidence representation of the initial defect segmentation result. The refined confidence representation is recombined across the entire image range to obtain a pixel confidence map that corresponds one-to-one with the detection area of ​​the composite film.

[0012] Optionally, the generation of the defect segmentation result specifically includes: The defect candidate region and the pixel confidence map are fed into the fusion processing unit. For each pixel, the candidate label and confidence value are read at the same time. The features are concatenated and convolution and nonlinear activation are performed in sequence to generate a joint feature map that comprehensively reflects the candidate information and the confidence information. The fusion features of each pixel in the fusion feature map are classified pixel by pixel to obtain the classification result of whether the current pixel belongs to the defect category or the background category. The result is compared with the preset classification threshold. When the defect response result is greater than or equal to the classification threshold, the pixel is marked as a defect pixel. Otherwise, it is marked as a background pixel, thus forming a pixel-level binary classification result map. Spatial connectivity and morphological processing are performed on the pixel-level binary classification result image. Adjacent defect pixels are clustered according to the preset neighborhood connection rules to obtain different independent defect regions. Dilation and erosion operations are performed on the boundaries of each defect region to remove isolated noise and compensate for local breaks, resulting in a set of defect regions with continuous contours and internal connectivity, which is output as the defect segmentation result of the composite film.

[0013] Optionally, obtaining the timing correction result specifically includes: The defect segmentation results are arranged in the order of acquisition to form a continuous sampling frame composed of multiple segmentation results. Temporal consistency constraints are applied to the continuous sampling frames. The translation displacement of any previous frame relative to the next frame in the running direction is calculated based on the running speed of the composite film on the production line and the camera trigger interval. The segmentation results of the previous frame are translated according to the translation displacement so that the translated segmentation results of the previous frame are aligned with the segmentation results of the next frame in spatial position. After completing the spatial alignment of the segmentation results of two adjacent frames, the segmentation categories of the previous frame segmentation result and the next frame segmentation result at the same pixel position are compared. The response change of each pixel position between two consecutive frames is calculated, and the current response change is compared with a preset temporal difference threshold to determine whether the segmentation result change of the pixel position in two consecutive frames exceeds the allowed time change range. The defect segmentation result of the next frame is corrected pixel by pixel based on the response change of each pixel position between two consecutive frames. When the response change of a certain pixel position is less than the temporal difference threshold, the segmentation category of the pixel position in the next frame is adjusted to be consistent with the segmentation category of the corresponding pixel in the previous frame after alignment. When the response change of a certain pixel position is greater than or equal to the temporal difference threshold, the segmentation category of the pixel position in the next frame remains unchanged. The correction operation is performed on all adjacent frames in the continuous frame sequence in sequence to obtain the temporal correction result.

[0014] Optionally, the determination of the thickness uniformity index and defect distribution characteristics specifically includes: The timing correction result is used as the defect segmentation input. The defect pixels are clustered according to the preset neighborhood connection rules to obtain different independent defect regions. The number of pixels in each defect region is calculated to obtain the defect area. The circumscribed rectangle is determined to obtain the start and end positions of the defect in the width direction and running direction of the composite film. The centroid position and aspect ratio of the defect region are recorded. All defect regions are statistically analyzed to obtain the defect distribution characteristics, including the number of defects, the total defect area, and the distribution of defects in the width direction and running direction. The thickness estimates of all pixel locations in the thickness estimation map are read, and the changes in the thickness estimates in the width direction and running direction of the composite film are calculated to obtain a thickness uniformity index that characterizes the degree of overall thickness fluctuation and local thickness variation.

[0015] Optionally, the determination of the composite film quality score specifically includes: Defect distribution characteristics and thickness uniformity index are used as inputs and combined in a preset order to form a quality assessment feature vector that characterizes the overall quality status of the composite film. According to the preset quality assessment weighting rules, weights are assigned to the defect-related features and thickness uniformity-related features in the quality assessment feature vector, respectively. The weighted operation of each feature and its corresponding weight is performed and summed to obtain a quality score value that represents the overall quality level of the composite film. The quality score value is then limited to a predetermined score range so that each roll or section of composite film corresponds to a unique quality score output. Based on the preset quality grade classification standard, the quality score value is compared with different grade thresholds. When the quality score value falls into different score ranges, the corresponding quality grade is determined. The quality score value and quality grade are recorded together as the quality assessment result of the composite film. The quality assessment result is sent to the production line control equipment to guide the adjustment of process parameters or trigger quality anomaly warnings.

[0016] The beneficial effects of this invention are: This invention achieves unified imaging of surface defects and internal structural anomalies in composite films by simultaneously acquiring transmission and reflection images on the composite film production line and combining geometric correction, brightness normalization, and spatial registration. This provides more comprehensive and stable visual information for subsequent defect identification and thickness analysis.

[0017] This invention employs a dual-modal defect localization module and a structure-thickness joint pixel confidence module based on LPC-Net to fuse transmission and reflection features across modes and introduce the thickness estimation map obtained from transmittance calibration into the pixel confidence modeling process. This significantly improves the segmentation accuracy of small-sized defects, low-contrast defects, and defects coupled with thickness anomalies, and effectively reduces missed detections and false detections under high background ratio conditions.

[0018] This invention ensures the continuity and traceability of defect appearance along the production line by performing motion compensation and temporal consistency constraints on the defect segmentation results of continuous sampling frames, thereby providing a reliable data foundation for subsequent defect statistics.

[0019] This invention constructs a quality assessment feature vector based on the acquisition of defect distribution characteristics and thickness uniformity indicators, and generates quality scores and quality grades through preset weighting rules, realizing quantitative and graded assessment of composite film quality. It can feed the quality assessment results back to the production line control equipment in real time to guide the adjustment of process parameters and trigger quality warnings, thereby improving the automation level and quality control capability of the composite film production process. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0021] Figure 1 This is a flowchart of a machine vision-based composite film quality assessment system proposed in this invention; Figure 2 This is a block diagram of the overall structure of a composite film quality assessment system based on machine vision proposed in this invention. Figure 3 This is a data flow diagram of a machine vision-based composite film quality assessment system proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figure 1-3 A machine vision-based composite film quality assessment system includes: The image acquisition module is used to simultaneously acquire the transmission and reflection images of the composite thin film and to complete the correction, normalization and registration processing. The feature extraction module is used to perform convolutional feature extraction on the preprocessed transmission and reflection images to generate transmission and reflection features. The defect localization module is used to input transmission and reflection features into the LPC-Net-based dual-modal defect localization module, perform feature fusion, and output defect candidate regions and initial defect segmentation results. The confidence generation module is used to obtain a thickness estimation map based on the transmittance calibration, and to jointly calculate and generate a pixel confidence map by combining the thickness estimation map with the initial defect segmentation results. The defect segmentation and temporal module is used to fuse defect candidate regions and pixel confidence maps to generate defect segmentation results, and to perform motion compensation and temporal correction on consecutive frames to obtain temporal correction results. The feature statistics module is used to count the number, area and distribution characteristics of defects based on the time-series correction results, and to calculate the thickness uniformity index based on the thickness estimation map. The quality assessment module is used to construct a quality assessment feature vector, generate a composite film quality score and quality grade, and send the results to the production line control equipment.

[0024] In this embodiment, the modules are interconnected using the following method: The transmission and reflection images of the composite thin film were acquired, and feature extraction was performed to obtain the transmission features and reflection features, respectively. Transmission and reflection features are input into the LPC-Net-based dual-modal defect localization module to perform cross-modal feature fusion, and defect candidate regions are determined based on the defect localization feature map to obtain the initial defect segmentation result. Transmittance calibration is performed on the transmission image to generate a thickness estimation map. The thickness estimation map and the initial defect segmentation result are simultaneously input into the structure-thickness joint pixel confidence module based on LPC-Net to generate a pixel confidence map. The defect candidate region is fused with the pixel confidence map to generate the defect segmentation result of the composite film; Temporal consistency constraints are applied to the continuous sampling frames of the composite film. The defect segmentation results of adjacent frames are matched with motion compensation according to the direction of production line movement. The defect segmentation results are corrected based on the pixel response differences of adjacent frames to obtain the temporal correction results. The defect distribution characteristics of the composite film are determined based on the time-series correction results, and the thickness uniformity index of the composite film is determined based on the thickness estimation diagram. The quality assessment feature vector is constructed based on the thickness uniformity index, and a composite film quality score is generated according to the preset weighting rules to determine the quality level of the composite film. The score is then sent to the production line control equipment for process adjustment and quality early warning.

[0025] In this embodiment, obtaining the transmission and reflection characteristics specifically includes: On both sides of the inspection area of ​​the composite film production line, a transmission imaging light source and a transmission industrial camera are respectively set up, and a reflection imaging light source and a reflection industrial camera. The transmission imaging light source and the transmission industrial camera are arranged opposite each other to collect the transmission image of the composite film. The reflection imaging light source and the reflection industrial camera are arranged with preset incident angle and observation angle to collect the reflection image of the composite film. The camera exposure time and trigger frequency are set according to the running speed of the composite film so that the transmission image and the reflection image are collected synchronously at the same physical position. Simultaneous acquisition of transmitted and reflected images, along with image preprocessing, including geometric distortion correction based on camera calibration parameters, normalization of overall brightness and contrast to reduce uneven illumination and noise, and spatial registration of transmitted and reflected images to ensure that the two images correspond to the same composite film region at pixel positions, resulting in preprocessed transmitted and reflected images. Feature extraction is performed on the preprocessed transmission and reflection images. Convolution, nonlinear activation and downsampling are performed sequentially to extract transmission and reflection features that characterize the internal structure and thickness variation of the composite film.

[0026] In this embodiment, obtaining the initial defect segmentation result specifically includes: Transmission and reflection features are input into a dual-modal defect localization module based on LPC-Net. Within the LPC-Net-based dual-modal defect localization module, different feature vectors at the same spatial location are weighted and summed according to preset weight parameters, and channel stitching is performed. Convolution and nonlinear activation operations are then executed sequentially to obtain a fused feature map that characterizes the defect response of the composite film. The LPC-Net is a two-stage industrial defect segmentation deep learning model consisting of a defect localization stage and a pixel confidence estimation stage, which is used to achieve high-precision defect segmentation in images with high background ratio and small defect area size. The fused feature map is input into the defect localization sub-network based on LPC-Net. Multi-layer convolution operation, nonlinear live operation and downsampling operation are performed in the defect localization sub-network in sequence to extract features and spatially compress the fused feature map, generating a defect localization feature map to characterize the defect response intensity at each position of the composite film. Based on the defect localization feature map, the initial segmentation probability of the current position belonging to the defect category is calculated for each spatial position to obtain the corresponding initial segmentation probability distribution. The initial segmentation probability distribution is subjected to threshold segmentation processing. A preset probability threshold is used to distinguish between defect pixels and background pixels. Positions with an initial segmentation probability greater than or equal to the probability threshold are marked as defect candidate pixels, and otherwise marked as background pixels. Defect candidate regions are generated based on the spatial connectivity of the defect candidate pixels. The defect candidate regions and the initial segmentation probability distribution are used together as the initial defect segmentation result for output.

[0027] This invention presents a dual-modal defect localization scheme based on LPC-Net. By weighted fusion and depth convolution extraction of transmission and reflection features at the same spatial location, surface defects and internal defects are simultaneously enhanced in the feature space, improving the response capability to small-sized, low-contrast defects. Furthermore, by outputting a pixel-level initial segmentation probability distribution and combining it with thresholds and spatial connectivity to generate defect candidate regions, the invention effectively suppresses background texture interference and isolated noise points, obtaining a clear and coherent initial defect segmentation result. This provides a high-quality input foundation for subsequent thickness prior introduction and pixel confidence refinement.

[0028] In this embodiment, the generation of the pixel confidence map specifically includes: Transmittance calibration is performed on the preprocessed transmission image. Different composite film calibration samples with known thicknesses are selected, and different calibration regions are divided. The actual thickness value and the average gray value in the transmission image are recorded for each calibration region. Based on the correspondence between the actual thickness value and the average gray value, a calibration relationship between transmission gray value and local thickness of composite film is established. The preprocessed transmission image of the composite film under test is input into the calibration relationship pixel by pixel. According to the calibration relationship, the gray value of each pixel is converted into the corresponding thickness estimate value. Thickness estimation is performed for all pixels of the entire image to generate a thickness estimation map that represents the local thickness distribution of composite film in space. The initial defect segmentation result and the thickness estimation map are input into the structure-thickness joint pixel confidence module based on LPC-Net. The structure-thickness joint pixel confidence module belongs to the pixel confidence sub-network in the two-stage network of LPC-Net. By combining the initial segmentation probability distribution with the thickness estimation value and local thickness change information at the pixel level, a joint feature vector including defect response, thickness estimation and neighborhood context is constructed for each position. The confidence level of the defect category to which each pixel belongs is re-estimated, forming a refined confidence representation of the initial defect segmentation result. The refined confidence representation is recombined across the entire image range to obtain a pixel confidence map that corresponds one-to-one with the detection area of ​​the composite film.

[0029] This invention obtains an accurate thickness estimation map by calibrating the transmission image, and jointly models the initial segmentation probability, thickness estimate, and local thickness changes in the structure-thickness joint pixel confidence module based on LPC-Net to generate a pixel-level confidence map. This enables targeted refinement and correction of the initial defect segmentation results, effectively distinguishes between normal areas caused by thickness fluctuations and real defect areas, and significantly improves the accuracy and robustness of composite film defect segmentation.

[0030] In this embodiment, the generation of the defect segmentation result specifically includes: The defect candidate region and the pixel confidence map are fed into the fusion processing unit. For each pixel, the candidate label and confidence value are read at the same time. The features are concatenated and convolution and nonlinear activation are performed in sequence to generate a joint feature map that comprehensively reflects the candidate information and the confidence information. The fusion features of each pixel in the fusion feature map are classified pixel by pixel to obtain the classification result of whether the current pixel belongs to the defect category or the background category. The result is compared with the preset classification threshold. When the defect response result is greater than or equal to the classification threshold, the pixel is marked as a defect pixel. Otherwise, it is marked as a background pixel, thus forming a pixel-level binary classification result map. Spatial connectivity and morphological processing are performed on the pixel-level binary classification result image. Adjacent defect pixels are clustered according to the preset neighborhood connection rules to obtain different independent defect regions. Dilation and erosion operations are performed on the boundaries of each defect region to remove isolated noise and compensate for local breaks, resulting in a set of defect regions with continuous contours and internal connectivity, which is output as the defect segmentation result of the composite film.

[0031] In this embodiment, obtaining the timing correction result specifically includes: The defect segmentation results are arranged in the order of acquisition to form a continuous sampling frame composed of multiple segmentation results. Temporal consistency constraints are applied to the continuous sampling frames. The translation displacement of any previous frame relative to the next frame in the running direction is calculated based on the running speed of the composite film on the production line and the camera trigger interval. The segmentation results of the previous frame are translated according to the translation displacement so that the translated segmentation results of the previous frame are aligned with the segmentation results of the next frame in spatial position. After completing the spatial alignment of the segmentation results of two adjacent frames, the segmentation categories of the previous frame segmentation result and the next frame segmentation result at the same pixel position are compared. The response change of each pixel position between two consecutive frames is calculated, and the current response change is compared with a preset temporal difference threshold to determine whether the segmentation result change of the pixel position in two consecutive frames exceeds the allowed time change range. The defect segmentation result of the next frame is corrected pixel by pixel based on the response change of each pixel position between two consecutive frames. When the response change of a certain pixel position is less than the temporal difference threshold, the segmentation category of the pixel position in the next frame is adjusted to be consistent with the segmentation category of the corresponding pixel in the previous frame after alignment. When the response change of a certain pixel position is greater than or equal to the temporal difference threshold, the segmentation category of the pixel position in the next frame remains unchanged. The correction operation is performed on all adjacent frames in the continuous frame sequence in sequence to obtain the temporal correction result.

[0032] This invention applies a temporal consistency constraint based on the production line running speed and camera trigger interval to the defect segmentation results arranged in the acquisition order. Based on motion compensation alignment of adjacent frames, the next frame segmentation result is corrected pixel by pixel according to the comparison result of pixel-level response change and temporal difference threshold. This eliminates random changes caused by noise, illumination fluctuations or local texture disturbances, avoids flickering or breakage of defect results between consecutive frames, and makes the defect appearance along the running direction of the composite film more continuous and stable. This is beneficial for the reliable tracking and statistical analysis of long strip defects and periodic defects.

[0033] In this embodiment, the determination of the thickness uniformity index and defect distribution characteristics specifically includes: The timing correction result is used as the defect segmentation input. The defect pixels are clustered according to the preset neighborhood connection rules to obtain different independent defect regions. The number of pixels in each defect region is calculated to obtain the defect area. The circumscribed rectangle is determined to obtain the start and end positions of the defect in the width direction and running direction of the composite film. The centroid position and aspect ratio of the defect region are recorded. All defect regions are statistically analyzed to obtain the defect distribution characteristics, including the number of defects, the total defect area, and the distribution of defects in the width direction and running direction. The thickness estimates of all pixel locations in the thickness estimation map are read, and the changes in the thickness estimates in the width direction and running direction of the composite film are calculated to obtain a thickness uniformity index that characterizes the degree of overall thickness fluctuation and local thickness variation.

[0034] In this embodiment, the determination of the composite film quality score specifically includes: Defect distribution characteristics and thickness uniformity index are used as inputs and combined in a preset order to form a quality assessment feature vector that characterizes the overall quality status of the composite film. According to the preset quality assessment weighting rules, weights are assigned to the defect-related features and thickness uniformity-related features in the quality assessment feature vector, respectively. The weighted operation of each feature and its corresponding weight is performed and summed to obtain a quality score value that represents the overall quality level of the composite film. The quality score value is then limited to a predetermined score range so that each roll or section of composite film corresponds to a unique quality score output. Based on the preset quality grade classification standard, the quality score value is compared with different grade thresholds. When the quality score value falls into different score ranges, the corresponding quality grade is determined. The quality score value and quality grade are recorded together as the quality assessment result of the composite film. The quality assessment result is sent to the production line control equipment to guide the adjustment of process parameters or trigger quality anomaly warnings.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to an industrial production line continuously producing high-barrier composite films. Online detection and quality assessment were performed on the surface defects, internal bubbles, and thickness uniformity of the entire roll of film. The production line was originally equipped with a traditional machine vision system using single-camera reflective imaging, which primarily relied on fixed threshold segmentation and connected component analysis for defect detection. It also included a separate online thickness gauge for random thickness measurements at key locations. In actual use, the original system exhibited a high rate of missed detection for small-sized, low-contrast defects, and the thickness measurement results were disconnected from the defect segmentation results, failing to provide a unified quality score and grade. This led to operators relying heavily on experience to adjust the process, resulting in significant quality fluctuations.

[0036] In this embodiment, the machine vision-based composite film quality assessment system of the present invention is integrated and upgraded with the existing production line. The image acquisition module arranges a reflective light source and a reflective industrial camera above and below the composite film, and a transmittance light source and a transmittance industrial camera, respectively, so that transmittance and reflective images are acquired simultaneously as the film passes through the inspection area. The system performs geometric distortion correction, brightness and contrast normalization, and spatial registration of transmittance / reflection images at the image acquisition end, ensuring that the same physical location corresponds one-to-one in the dual-modal images. The feature extraction module performs convolutional feature extraction on the preprocessed transmittance and reflection images, generating transmittance features characterizing internal structure and thickness variations, and reflectance features characterizing surface texture and scratches, respectively.

[0037] In the defect analysis phase, the defect localization module inputs transmission and reflection features into a dual-modal defect localization network based on LPC-Net. Dual-modal features at the same spatial location are fused according to preset weights, and the defect response is extracted through multi-layer convolution and nonlinear activation, outputting pixel-level initial segmentation probabilities and candidate defect regions. Subsequently, the confidence generation module first uses a transmittance-thickness mapping relationship established from calibration samples of different thicknesses to convert the grayscale of the transmission image pixel by pixel into a thickness estimate, forming a thickness estimation map. Then, the initial defect segmentation result and the thickness estimation map are input into a structure-thickness joint pixel confidence network. At the pixel level, the defect response, thickness estimation, and local thickness changes are jointly modeled to obtain a pixel confidence map, refining the initial segmentation result. The defect segmentation and temporal sequence module fuses the candidate defect regions with the pixel confidence map at the pixel level to generate a single-frame defect segmentation result. Combining the film running speed and camera trigger interval, motion compensation and temporal consistency constraints are performed on continuously sampled frames. Pixels with response changes not exceeding a threshold are subject to category correction to obtain a continuous and consistent temporal correction result along the running direction.

[0038] During the quality assessment phase, the feature statistics module, based on the time-corrected defect segmentation results, statistically analyzes the defect distribution characteristics of each roll of composite film, including the number of defects, total defect area, defect density per unit area, defect distribution range in the width and running directions, aspect ratio and center of gravity of typical defect areas. Simultaneously, it calculates thickness uniformity indicators such as average thickness, thickness range, thickness standard deviation, and thickness fluctuation amplitude in the width and running directions on the thickness estimation map. The quality assessment module combines these defect and thickness characteristics into a quality assessment feature vector. Based on a weighting rule pre-determined by process experts and historical good product data, it generates a quality score ranging from 0 to 100, classifies the quality into grades according to the score range, and sends the quality score and grade results to the production line control system in real time. Operators can adjust process parameters such as coating amount, lamination pressure, and cooling rate based on the trend changes in the quality score. When the score falls below the warning threshold, an automatic alarm or roll marking is triggered.

[0039] To verify the effectiveness of the system of the present invention, during the continuous production of multiple batches of composite films on the production line, a statistical analysis was performed on samples from the same batch, comparing the traditional single-modal vision scheme, the system of the present invention, and a convolutional neural network segmentation scheme based solely on single-reflection images. The specific comparison results are shown in Table 1. Table 1. Performance Comparison of Different Detection Schemes in Online Quality Assessment of Composite Thin Films

[0040] As shown in Table 1, the system of this embodiment outperforms existing vision solutions A and B in all key indicators: the recall rate for small-sized defects is increased to 96.7%, the overall defect false negative rate is reduced to 1.6%, and the number of false alarms per unit area is reduced to approximately 3.2 per 1000㎡; the correlation coefficient between the online calculated thickness uniformity score and the offline measurement results reaches 0.96, with an average relative error of only 1.9%, indicating that the thickness assessment results are closer to the true value; in terms of early warning of non-conforming roll materials, the system of this embodiment achieves a warning accuracy rate of 97.2%, significantly improving defect detection accuracy and quality assessment reliability while ensuring that the detection frame rate meets the production cycle.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine vision-based composite film quality evaluation system, characterized in that, The method comprises the following steps: An image acquisition module is used to synchronously acquire the transmission image and the reflection image of the composite film, and to complete correction, normalization and registration processing; A feature extraction module is used to perform convolution feature extraction on the preprocessed transmission image and reflection image, to generate transmission features and reflection features; A defect positioning module is used to input the transmission features and the reflection features into the dual-modal defect positioning module based on LPC-Net, to perform feature fusion and output defect candidate regions and defect initial segmentation results; A confidence generation module is used to obtain a thickness estimation map based on the transmittance calibration, and to jointly calculate the thickness estimation map and the defect initial segmentation results to generate a pixel confidence map; A defect segmentation and timing module is used to fuse the defect candidate regions and the pixel confidence map to generate defect segmentation results, and to perform motion compensation and timing correction on the continuous frames to obtain timing correction results; A feature statistics module is used to count the defect number, area and distribution characteristics based on the timing correction results, and to calculate the thickness uniformity index based on the thickness estimation map; A quality evaluation module is used to construct a quality evaluation feature vector, to generate a composite film quality score and a quality grade, and to send the results to a production line control device.

2. The machine vision-based composite film quality evaluation system of claim 1, wherein, The modules are realized through the following methods: The transmission image and the reflection image of the composite film are acquired, and the transmission features and the reflection features are obtained through feature extraction; The transmission features and the reflection features are input into the dual-modal defect positioning module based on LPC-Net, cross-modal feature fusion is performed, and defect candidate regions are determined according to the defect positioning feature map to obtain defect initial segmentation results; Transmittance calibration is performed on the transmission image to generate a thickness estimation map, which is input into the structure-thickness joint pixel confidence module based on LPC-Net together with the defect initial segmentation results to generate a pixel confidence map; The defect candidate regions and the pixel confidence map are fused to generate defect segmentation results; The continuous sampling frames of the composite film are subjected to timing consistency constraints, the defect segmentation results of adjacent frames are motion compensated and matched according to the production line motion direction, and the defect segmentation results are corrected based on the pixel response difference of adjacent frames to obtain timing correction results; The defect distribution characteristics of the composite film are determined according to the timing correction results, and the thickness uniformity index of the composite film is determined according to the thickness estimation map; The thickness uniformity index is used to construct a quality evaluation feature vector, a composite film quality score is generated according to a pre-set weighting rule, the quality grade of the composite film is determined, and the results are sent to a production line control device.

3. The machine vision-based composite film quality evaluation system of claim 2, wherein, The transmission features and the reflection features are obtained as follows: Transmission imaging light sources and transmission industrial cameras, reflection imaging light sources and reflection industrial cameras are arranged on both sides of the detection area of the composite film production line, and the camera exposure time and trigger frequency are set according to the running speed of the composite film; The transmission image and the reflection image are synchronously acquired, and image preprocessing is performed simultaneously, including geometric distortion correction according to the camera calibration parameters, normalization of overall brightness and contrast, and spatial registration of the transmission image and the reflection image to obtain preprocessed transmission image and reflection image; The transmission image and the reflection image after preprocessing are subjected to feature extraction, convolution operation, nonlinear activation and down-sampling in sequence to extract transmission features and reflection features representing internal structure and thickness variation of the composite film.

4. The machine vision-based composite film quality evaluation system of claim 2, wherein, The obtaining of the initial segmentation result of the defect specifically includes: The transmission features and the reflection features are input into a dual-modal defect positioning module based on LPC-Net, in which different feature vectors at the same spatial position are subjected to weighted addition and channel splicing processing according to preset corresponding weight parameters, and convolution operation and nonlinear activation operation are sequentially performed to obtain a fusion feature map; The fusion feature map is input into a defect positioning subnetwork based on LPC-Net, in which multi-layer convolution operation, nonlinear activation operation and down-sampling operation are sequentially performed to extract features and compress space of the fusion feature map, generate a defect positioning feature map, and calculate an initial segmentation probability of each spatial position belonging to a defect category based on the defect positioning feature map to obtain a corresponding initial segmentation probability distribution; The initial segmentation probability distribution is subjected to threshold segmentation processing, a preset probability threshold is set, positions with initial segmentation probabilities greater than or equal to the probability threshold are marked as defect candidate pixels, otherwise, they are marked as background pixels, and defect candidate regions are generated according to the spatial connection relationship of the defect candidate pixels, which are output as the initial segmentation result of the defect together with the initial segmentation probability distribution.

5. The machine vision-based composite film quality evaluation system of claim 2, wherein, The generation of the pixel confidence map specifically includes: The transmission image after preprocessing is subjected to transmittance calibration, composite film calibration samples are selected, different calibration regions are divided, actual thickness values and average gray values in the transmission image corresponding to each calibration region are recorded, a calibration relationship between transmission gray and local thickness of the composite film is established based on the corresponding relationship between the actual thickness values and the average gray values, and the transmission image after preprocessing of the composite film to be measured is input into the calibration relationship pixel by pixel, the gray value of each pixel is converted into a corresponding thickness estimation value according to the calibration relationship, thickness estimation of all pixels of the entire image is completed, and a thickness estimation map is generated; The initial segmentation result of the defect and the thickness estimation map are input into a structure-thickness joint pixel confidence module based on LPC-Net, the initial segmentation probability distribution is combined with the thickness estimation value and the local thickness variation information at the same position at the pixel level, a joint feature vector is constructed for each position, the confidence degree of the defect category to which each pixel belongs is re-estimated, and a refined confidence representation is formed; The refined confidence representation is reorganized in the entire image range to obtain a pixel confidence map corresponding to the composite film detection region one by one.

6. The machine vision-based composite film quality evaluation system of claim 2, wherein, The generation of the defect segmentation result specifically includes: The defect candidate region and the pixel confidence map are input into a fusion processing unit, the candidate mark and the confidence value of each pixel are read at the same time, feature splicing is performed, convolution operation and nonlinear activation are sequentially performed, and a joint feature map is generated; The fusion features of each pixel in the fusion feature map are classified pixel by pixel to obtain a classification result, and the classification result is compared with a preset classification threshold. When the defect response result is greater than or equal to the classification threshold, the pixel is marked as a defect pixel, otherwise, the pixel is marked as a background pixel, thereby forming a pixel-level binary classification result map; The pixel-level binary classification result map is subjected to spatial connectivity and morphological processing. According to a preset neighborhood connection rule, adjacent defect pixels are clustered in a connected domain to obtain different independent defect regions. The boundaries of each defect region are subjected to dilation and erosion operations to remove isolated noise and compensate for local breaks, thereby obtaining a defect region set with continuous contours and internal connectivity, which is output as the defect segmentation result of the composite film.

7. The machine vision-based composite film quality evaluation system of claim 2, wherein, The time sequence correction result is obtained specifically as follows: The defect segmentation results are arranged according to the acquisition sequence to form a continuous sampling frame composed of multiple segmentation results. The continuous sampling frame is subjected to time sequence consistency constraint. According to the running speed of the composite film on the production line and the camera trigger interval, the translation displacement of any previous frame relative to the next frame in the running direction is calculated, and the previous frame segmentation result is subjected to translation processing according to the translation displacement. After the spatial alignment of the adjacent two frame segmentation results is completed, the segmentation categories at the same pixel position of the aligned previous frame segmentation result and the next frame segmentation result are compared. The response change of each pixel position between the two continuous frames is calculated, and the current response change is compared with a preset time sequence difference threshold. The defect segmentation result of the next frame is corrected pixel by pixel according to the response change of each pixel position between the two continuous frames. The correction operation is sequentially performed on all adjacent frames in the continuous frame sequence to obtain the time sequence correction result.

8. The machine vision-based composite film quality evaluation system of claim 2, wherein, The thickness uniformity index and the defect distribution feature are determined specifically as follows: The time sequence correction result is taken as the defect segmentation input. According to a preset neighborhood connection rule, the defect pixels are clustered to obtain different independent defect regions. The number of pixels in each defect region is calculated to obtain the defect area. The circumscribed rectangle is determined to obtain the start and end positions of the defect in the width direction and the running direction of the composite film. The center of gravity position and the aspect ratio of the defect region are recorded. All defect regions are counted to obtain the defect distribution feature. The thickness estimation values of all pixel positions in the thickness estimation map are read, and the changes of the thickness estimation values in the width direction and the running direction of the composite film are calculated to obtain the thickness uniformity index.

9. The machine vision-based composite film quality evaluation system of claim 2, wherein, The determination of the composite film quality score specifically includes: The defect distribution feature and the thickness uniformity index are taken as inputs to form a quality evaluation feature vector in a preset order; According to a preset quality evaluation weighting rule, the defect-related features and the thickness uniformity-related features in the quality evaluation feature vector are assigned weights. The features and the corresponding weights are subjected to weighted operation and summation to obtain a quality score value, and the quality score value is limited in a predetermined score interval range. According to the preset quality grade division standard, the quality score value is compared with different grade threshold values, when the quality score value falls into different score intervals, the corresponding quality grade is determined respectively, the quality score value and the quality grade are recorded together as the quality evaluation result of the composite film, and the quality evaluation result is sent to the production line control device.

Images