Fabric color difference online detection system based on multispectral imaging and convolutional neural network

Abstract

The application relates to a fabric color difference online detection system based on multispectral imaging and a convolutional neural network, which comprises a multispectral imaging module, an image preprocessing and illumination correction module, a CNN color difference intelligent identification and calculation module and a feedback control and execution module. A high-precision spectral calibration unit is first set, multispectral imaging module is used to collect multi-channel image data of the fabric under visible light and near-infrared wave bands, the image data is processed, a lightweight convolutional neural network embedded with an attention mechanism is adopted to extract features and identify color difference regions of the preprocessed image, the CNN model adopts a deformable convolution kernel to adapt to the texture direction of different fabrics, the identified color difference grade and position information are fed back to a dyeing machine control system in real time, and process parameters are adjusted. The application solves the problems of low recognition accuracy and poor real-time performance of traditional manual visual detection and existing machine vision methods under complex textures and illumination changes.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and quality inspection technology in textile printing and dyeing, specifically to an online fabric color difference detection system based on multispectral imaging and convolutional neural networks. Background Technology

[0002] Textiles are indispensable materials in people's lives, and textile quality is the core of maintaining industrial competitiveness. Among numerous quality indicators, color difference is one of the most direct and critical factors in evaluating the appearance quality of textiles. Achieving efficient and accurate color difference detection and control has long been a core demand and technical challenge for the textile printing and dyeing industry.

[0003] Color difference in fabrics is a key factor affecting the appearance quality and commercial value of textiles. Currently, color difference detection in industrial applications mainly relies on visual inspection, instrumental testing, spectral analysis, or methods based on... Inspection methods using machine vision.

[0004] Manual visual inspection is susceptible to the influence of human experience, physiological state, and psychological factors, resulting in high subjectivity, low efficiency, and, most importantly, difficulty in quantifying standards. While instrumental measurement provides objective data and overcomes the subjectivity issue, it is mostly contact-based and single-point measurement, limiting it to sampling inspection of produced fabrics and exhibiting slow measurement speed. Fiber optic spectrometers can provide continuous spectral data; however, traditional spectrometers can only perform single-point measurements, losing spatial information, failing to locate color difference areas, and unable to assess color uniformity. Machine vision inspection methods utilize computer vision for automatic inspection, but they are extremely sensitive to changes in ambient lighting. Complex textures and fuzzy patterns on the fabric surface can severely interfere with the accurate extraction of color information. Cameras cannot distinguish between objects with different spectral reflectance curves that appear to the human eye as the same color under a specific light source. This results in fundamental errors in color matching and accurate color difference judgment, limiting the upper limit of detection accuracy.

[0005] More critically, the existing technologies suffer from a severe disconnect between detection and control, making effective online feedback and adjustment difficult. The results of manual visual inspection and instrument sampling are significantly delayed; by the time color differences are detected, a large number of defective products have already been generated, and feedback and adjustment are too late. Even existing machine vision systems that can identify color differences mostly remain at the alarm or classification stage, lacking a closed-loop control link deeply integrated with production line actuators. Their feedback and adjustment mechanisms are often simplistic and crude, typically making unidirectional and lagging adjustments based solely on the overall average color difference. They cannot utilize key information such as the spatial distribution and gradient changes of color differences to coordinate, accurately, and proactively control multiple process parameters such as dye injection, temperature, and machine speed. This open-loop detection and offline decision-making model leads to passive and lagging quality control, failing to suppress color differences at their source.

[0006] Current technologies present a dilemma: those capable of full-field imaging lack sufficient precision and are susceptible to interference, while high-precision technologies cannot achieve full-field imaging and online detection. Furthermore, existing technological systems generally lack closed-loop control capabilities from precise perception to intelligent decision-making and real-time execution. To address these technical issues, it is indeed necessary to provide a system that combines the high precision of spectral analysis with the full-field, high-speed advantages of machine vision, and ultimately deeply integrate it with the production execution system to form a complete quality control closed loop from precise perception to intelligent decision-making and real-time regulation, thereby completely resolving the long-standing pain point of color difference control in the textile printing and dyeing industry. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an online intelligent detection and closed-loop control system and method for fabric color difference based on the fusion of multispectral imaging and spectral analysis. This system aims to solve the problems of traditional detection methods, such as high subjectivity, low efficiency, insufficient accuracy, and inability to achieve online closed-loop control. By integrating multispectral imaging, spectral analysis, deep learning, and automatic control technologies, it achieves full-process automation from accurate color perception to intelligent process control.

[0008] To achieve the aforementioned first objective, the technical solution adopted by the present invention is an online fabric color difference detection system and method based on multispectral imaging and convolutional neural networks, characterized in that it includes a multispectral imaging module, an image preprocessing and illumination correction module, a CNN color difference intelligent recognition and calculation module, and a feedback control and execution module.

[0009] The multispectral imaging module uses a multispectral industrial camera and a standardized lighting system to acquire image data of fabrics in operation under multiple specific spectral bands such as visible light and near-infrared light, and transmits the multispectral image data to the edge computing server in real time using industrial Ethernet and communication networks. By acquiring spectral dimension information that surpasses the traditional RGB three-channel, this module provides a fundamental data foundation for overcoming the metamerism problem and achieving high-precision color difference detection.

[0010] The image preprocessing and illumination correction module uses a bilateral filtering algorithm and a deep learning-based illumination estimation algorithm to process the raw image data acquired by the multispectral imaging module. The bilateral filtering effectively suppresses image noise while perfectly preserving the edge information of the fabric color. The illumination estimation based on a convolutional neural network can achieve color constancy correction of the imaging environment, eliminate color deviations introduced by changes in ambient light, and ensure that the image data input to subsequent modules is true and reliable in color.

[0011] The CNN color difference intelligent recognition and calculation module adopts a lightweight convolutional neural network architecture with embedded attention mechanism and deformable convolution to perform deep feature extraction and color difference region identification on preprocessed multispectral images. This module integrates principal component analysis-extreme learning machine model to achieve accurate conversion from multispectral data to CIE Lab color space, and integrates the international textile industry standard CMC color difference formula, finally outputting a quantized color difference value ΔE that is highly consistent with visual perception. This module focuses on key color features through attention mechanism and adapts to changes in fabric texture through deformable convolution, thereby meeting the high frame rate requirements of industrial online inspection while ensuring high accuracy and recall.

[0012] The feedback control and execution module compares the color difference value ΔE output by the CNN color difference intelligent recognition and calculation module with the preset ISO quality threshold in real time. When the color difference is abnormal, it sends control commands to the PLC control system of the dyeing machine through the OPC UA industrial communication protocol to drive the dye liquor regulating valve, temperature controller, speed regulator and other actuators to realize real-time, dynamic and closed-loop adjustment of dyeing process parameters, thereby actively correcting the color difference from the source and completing a fully automatic quality control closed loop from perception to decision to execution.

[0013] The multispectral imaging and convolutional neural network-based online fabric color difference detection system of the present invention further comprises: the multispectral image acquisition unit includes a multispectral industrial camera, a fiber optic spectrometer, a standard light source box, a colorimeter, a high-resolution scanner, and a near-infrared spectrometer.

[0014] The multispectral imaging and convolutional neural network-based online fabric color difference detection system of the present invention further comprises: the image data acquired by the multispectral imaging module includes visible light RGB images, near-infrared reflectance images, ultraviolet fluorescence images, and thermal infrared distribution images of the fabric; the spectral data acquired by the spectral analysis module includes 400-700nm visible spectral reflectance curves, 700-1000nm near-infrared spectral characteristic peaks, and spectral reflectance ratios of specific bands; the quantitative indicators output by the color difference calculation module include CIELAB color space ΔE value, CMC color difference value, CIE94 color difference metric, and CIEDE2000 comprehensive color difference evaluation value.

[0015] The multispectral imaging and convolutional neural network-based online fabric color difference detection system of the present invention further includes the following color difference detection and evaluation parameters: the overall color difference ΔE of the fabric, the local color difference distribution map, the area ratio of the color difference region, the location coordinates of the maximum color difference region, the color difference gradient change trend, the metamerism index, the color difference grade assessment under the standard light source, and the color difference pass rate under the customer-specified light source conditions.

[0016] The multispectral imaging and convolutional neural network-based online fabric color difference detection system of the present invention further includes: the training data sources of the color space accurate conversion module include multispectral image samples of standard color cards, standard spectral reflectance data measured by fiber optic spectrometers, a database of color characteristics of qualified samples in historical production, color sample data under multi-light source conditions generated by digital simulation, and visual evaluation results of color difference levels by color experts.

[0017] The online fabric color difference detection system based on multispectral imaging and convolutional neural networks of the present invention further comprises the following steps: the operation of the color space accurate conversion module includes the following process: constructing p spectral feature vector matrices based on p sets of standard color card spectral data and p color feature vector matrices based on p sets of standard color feature values ​​as training samples input to the already trained principal component analysis-extreme learning machine model; the trained model will derive the nonlinear mapping relationship between each spectral feature and color feature value.

[0018] The multispectral imaging and convolutional neural network-based online fabric color difference detection system of the present invention further comprises: the color difference intelligent recognition and calculation module is constructed on the basis of the color space accurate conversion module, and by inputting a set of multispectral image data, combined with the color feature values ​​formed by the color space conversion module, the color difference value between the fabric and the standard sample is calculated in real time using a lightweight convolutional neural network model;

[0019] The system uses the following performance indicators to quantitatively evaluate and optimize the color difference detection effect:

[0020] Accuracy ;

[0021] Recall rate ;

[0022] Mean Precision ;

[0023] in, This represents the number of positive samples that were correctly identified. This represents the number of negative samples that were misidentified. This represents the number of positive samples that were incorrectly identified. This represents the total number of categories. This is the sequence number of the current detection category. For the first Average precision for each category;

[0024] The detection speed of the color difference intelligent recognition and calculation module in the multispectral imaging and convolutional neural network-based online fabric color difference detection system of the present invention is measured by the frame rate, and its calculation formula is as follows:

[0025] Frame rate

[0026] in, The total number of images processed within a specified time period. The total time consumed in processing these images; when the system is deployed on an edge computing device, its The frame rate should be no less than 30 frames per second to meet the real-time requirements of high-speed online detection in industrial production lines.

[0027] The multispectral imaging and convolutional neural network-based online fabric color difference detection system of the present invention further includes: the feedback control and execution module further includes a process knowledge base, which stores process adjustment strategies corresponding to different fabric materials, dye types and color difference levels, supports intelligent decision-making based on rules and cases, and realizes self-learning and optimization of color difference closed-loop control.

[0028] This invention provides an online fabric color difference detection system based on multispectral imaging and convolutional neural networks. It offers the following advantages:

[0029] 1. This invention, through multispectral imaging and spectral analysis technology, can acquire richer spectral information and reduce ambient light interference and subjective misjudgment compared to traditional manual visual methods and RGB vision methods.

[0030] 2. By constructing a feedback mechanism from detection to regulation, this invention can monitor color difference online and automatically adjust process parameters, which helps to reduce the generation of batch defective products and reduce raw material waste.

[0031] 3. This invention realizes the automated operation from image acquisition to parameter adjustment, reduces the reliance on manual quality inspection, and meets the needs of continuous operation of industrial production lines.

[0032] 4. This invention introduces a channel attention mechanism and deformable convolution, enabling the model to adapt to the texture changes of different fabrics and maintain good detection performance in complex backgrounds. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the online fabric color difference detection system of the present invention;

[0034] Figure 2 This is an operational diagram of the online fabric color difference detection system of the present invention. Detailed Implementation

[0035] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see the appendix Figure 1 - Appendix Figure 2 This invention provides an online fabric color difference detection system based on multispectral imaging and convolutional neural networks. The system mainly consists of a multispectral imaging module, an image preprocessing and illumination correction module, a CNN color difference intelligent recognition and calculation module, and a feedback control and execution module. The modules interact via industrial Ethernet, which employs Time-Sensitive Networking (TSN) technology and uses a traffic shaping mechanism to ensure real-time data transmission.

[0037] The multispectral imaging module is equipped with a high-precision spectral calibration unit, which specifically includes a fiber optic spectrometer and a standard white board. The fiber optic spectrometer is used to monitor the spectral distribution of the light source in real time and to collect spectral reflectance data from the standard white board and standard color card. The multispectral imaging module is also equipped with a composite illumination system, which includes a visible light source (using a D65 standard light source to simulate natural light), a near-infrared supplemental light (wavelength coverage 700-1000nm), and an ultraviolet excitation source (center wavelength 365nm). With the cooperation of the above composite light source, the multispectral industrial camera simultaneously captures multi-channel image data of the fabric under different lighting conditions. Specifically, the acquired data includes visible light RGB images, near-infrared reflectance images, and ultraviolet fluorescence images under ultraviolet light excitation. This multi-band acquisition method can obtain the response characteristics of the fabric under different spectra, thereby identifying metamerism phenomena that are difficult to distinguish under a single light source.

[0038] The image preprocessing and illumination correction module receives raw data from the imaging module. To eliminate the impact of ambient light fluctuations and equipment noise on image quality in industrial settings, this module first performs noise reduction on the image. A bilateral filtering algorithm is used here, which can filter out high-frequency noise while preserving the gradient information of the fabric texture edges, avoiding image blurring. Subsequently, illumination correction is performed. The system utilizes a color constancy model based on a convolutional neural network to learn image features under different lighting conditions, estimate and remove the illumination components in the scene, thereby restoring the true color of the fabric under standard light sources.

[0039] After basic preprocessing, to support subsequent color difference calculations, the multispectral image data needs to be mapped to the standard CIE Lab color space. In this embodiment, data obtained from a high-precision spectral calibration unit is used as a benchmark to construct a mapping model from multispectral data to the standard color space. Specifically, principal component analysis (PCA) combined with extreme learning machine (ELM) is employed.

[0040] The construction and operation process of this mapping model includes:

[0041] Step S101: Obtain the multispectral image data of the standard color card and the corresponding standard spectral reflectance data.

[0042] Step S102: Construct p spectral feature vector matrices based on the spectral data of p sets of standard color cards, and p color feature vector matrices based on the color feature values ​​of p sets of standard colors.

[0043] Step S103: Input the above matrix as a training sample into the Principal Component Analysis-Extreme Learning Machine model. The model learns through training and establishes a nonlinear mapping relationship between spectral features and color feature values.

[0044] Step S104: In the online detection stage, the real-time acquired multispectral image data is input into the trained model, and the corresponding CIE Lab color space values ​​are output.

[0045] The CNN color difference intelligent recognition and calculation module is the core processing unit of the system. This module is built on a lightweight convolutional neural network (CNN). To adapt to the complex textures of different fabrics (such as twill and satin), a deformable convolution kernel is used in the network structure. The deformable convolution kernel adds an extra offset to the standard convolution sampling points, allowing its sampling grid to deform freely according to the shape of the target, thereby extracting the color features of the fabric surface more accurately without being affected by the texture geometry. In addition, a channel attention mechanism is embedded in the network to weight the features of different spectral channels, automatically enhancing the response to color difference-sensitive channels (such as specific near-infrared band channels) and suppressing interference from invalid background information.

[0046] This module identifies color difference regions based on extracted features and calculates the quantified color difference value ΔE using the international standard CMC (l:c) color difference formula. Compared to simple Euclidean distance, the CMC color difference formula considers the differences in human eye sensitivity to different color regions, resulting in evaluation results that better align with human visual perception. In its calculation formula, 'c' represents the luminance weight, and 'c' represents the chroma weight. In this embodiment, for the application scenario of textile detection, the following is set... This means that the CMC (2:1) standard is used for calculation. In order to evaluate the color difference level, this module will also compare the calculated Delta E value with the gray scale level in international standards (such as ISO 105-A02), and divide the color difference into 1-5 levels (where level 5 is no color difference and level 1 is severe color difference), thereby outputting specific color difference level information.

[0047] The system uses the following performance indicators to quantitatively evaluate the actual effect of the color difference detection model: accuracy. The calculation formula is:

[0048] ;

[0049] Recall rate The calculation formula is:

[0050] ;

[0051] Mean Precision The calculation formula is:

[0052] ;

[0053] in, This represents the number of positive samples (areas with color difference) that were correctly identified. This represents the number of negative samples (normal areas) that were incorrectly identified as color difference areas. This represents the number of positive samples that were incorrectly identified as normal regions. This represents the total number of categories. This is the sequence number of the current detection category. For the first Average precision for each category.

[0054] Furthermore, to meet the speed requirements of industrial production lines, the system's detection speed is measured by the frame rate, calculated using the following formula:

[0055] ;

[0056] in, The total number of images processed within a specified time period. The total time consumed in processing these images. In this embodiment, the system is deployed on an edge computing device, and after lightweight pruning and model quantization optimization, its... Maintain a frame rate of 30 frames per second or higher.

[0057] The feedback control and execution module is used to receive the identified color difference level and color difference value. The module receives the location information and sends it to the dyeing machine control system for process parameter adjustment. The module has a preset color difference threshold (e.g., a tolerance range based on ISO standards). When the detected color difference value... When E exceeds the preset threshold, the module activates the adjustment mechanism.

[0058] The specific implementation steps for feedback adjustment are as follows:

[0059] Step S201: Read the current color difference value E. Color bias (e.g., reddish or bluish) and the specific location of the color difference (left, center, or right).

[0060] Step S202: Match the corresponding adjustment strategy based on the rules in the process knowledge base. The process knowledge base stores the association model between different fabric materials, dye types, and process parameters.

[0061] Step S203: Generate control commands. Send specific adjustment parameters to the PLC control system of the dyeing machine via the OPC UA industrial communication protocol.

[0062] Step S204: The dyeing machine actuator responds to the command. For example, if the overall color depth is detected to be too light, the dye injection rate is increased; if a color difference is detected between the left and right sides of the fabric, a pressure adjustment command is sent to the roller control unit or the temperature of the left and right temperature zones is adjusted; if uneven dyeing is detected, the fabric conveying speed is appropriately reduced to increase the dyeing time.

[0063] Step S205: This module also includes a process parameter adaptive adjustment unit, which uses a color difference trend prediction model and historical data to predict color difference changes in the future, performs feedforward control, and fine-tunes parameters in advance to suppress the generation of color difference.

[0064] For the industrial camera selection, the specific programming implementation of the OPC UA protocol stack, and the conventional bilateral filtering algorithm code implementation involved in the above embodiments, those skilled in the art can refer to relevant technical manuals or open-source resources, as these are well-known technologies in the field and will not be elaborated further here. Through the collaborative work of the above modules, this system realizes a fully automated process from spectral image acquisition and high-precision color analysis to closed-loop control of the dyeing process.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An online fabric color difference detection system based on multispectral imaging and convolutional neural networks, characterized in that: It includes a multispectral imaging module, an image preprocessing and illumination correction module, a CNN color difference intelligent recognition and calculation module, and a feedback control and execution module; The multispectral imaging module is equipped with a high-precision spectral calibration unit for acquiring multi-channel image data of the fabric in the visible and near-infrared bands; the image preprocessing and illumination correction module is used to perform noise reduction and illumination correction processing on the image data; the CNN color difference intelligent recognition and calculation module uses a convolutional neural network with an embedded attention mechanism to extract features and identify color difference regions in the preprocessed image, and calculate the color difference value; the feedback control and execution module is used to receive the identified color difference level and location information, and send it to the dyeing machine control system for process parameter adjustment.

2. The online fabric color difference detection system based on multispectral imaging and convolutional neural network according to claim 1, characterized in that, The high-precision spectral calibration unit includes a fiber optic spectrometer, used to establish a benchmark mapping relationship between the spectral reflectance of the standard color card and the color feature values; the multispectral imaging module is configured to control a multispectral camera to simultaneously capture and acquire image data of the fabric in multiple specific spectral bands.

3. The online fabric color difference detection system based on multispectral imaging and convolutional neural network according to claim 1, characterized in that, The color difference detection and evaluation parameters include the overall color difference ΔE of the fabric, the local color difference distribution map, the area ratio of the color difference region, the location coordinates of the maximum color difference region, the color difference gradient change trend, the metamerism index, the color difference grade assessment under standard light source, and the color difference pass rate under customer-specified light source conditions.

4. The online fabric color difference detection system based on multispectral imaging and convolutional neural network according to claim 1, characterized in that, The training data sources for the color space accurate conversion module include multispectral image samples of standard color cards, standard spectral reflectance data measured by fiber optic spectrometers, a database of color characteristics of qualified samples from historical production, color sample data under multi-light source conditions generated through digital simulation, and visual evaluation results of color difference levels by color experts.

5. The online fabric color difference detection system based on multispectral imaging and convolutional neural network according to claim 1, characterized in that, The operation of the color space accurate conversion module includes the following process: constructing p spectral feature vector matrices based on the p sets of standard color card spectral data and p color feature vector matrices based on the p sets of standard color feature values ​​as training samples input to the already trained principal component analysis-extreme learning machine model; the trained model will derive the nonlinear mapping relationship between each spectral feature and color feature value.

6. The online fabric color difference detection system based on multispectral imaging and convolutional neural network according to claim 1, characterized in that, The color difference intelligent recognition and calculation module is built on the basis of the color space accurate conversion module. By inputting a set of multispectral image data and combining the color feature values ​​formed by the color space conversion module, the module uses a lightweight convolutional neural network model to calculate the color difference between the fabric and the standard sample in real time. The system uses the following performance indicators to quantitatively evaluate and optimize the color difference detection effect: Accuracy ; Recall rate ; Mean accuracy ; in, This represents the number of positive samples that were correctly identified. This represents the number of negative samples that were misidentified. This represents the number of positive samples that were incorrectly identified. This represents the total number of categories. This is the sequence number of the current detection category. For the first Average precision for each category.

7. The online fabric color difference detection system based on multispectral imaging and convolutional neural network according to claim 1, characterized in that, The detection speed of the color difference intelligent recognition and calculation module is measured by the frame rate, and its calculation formula is as follows: Frame rate ; in, The total number of images processed within a specified time period. The total time consumed in processing these images; when the system is deployed on an edge computing device, its The frame rate should be no less than 30 frames per second to meet the real-time requirements of high-speed online detection in industrial production lines.

8. The online fabric color difference detection system based on multispectral imaging and convolutional neural network according to claim 1, characterized in that, The feedback control and execution module further includes a process knowledge base, which stores process adjustment strategies corresponding to different fabric materials, dye types and color difference levels, supports intelligent decision-making based on rules and cases, and realizes self-learning and optimization of color difference closed-loop control.

Images