CN-121981940-A - Intelligent screening and quantitative evaluation method and device for ancient textile diseases

CN121981940ACN 121981940 ACN121981940 ACN 121981940ACN-121981940-A

Abstract

The application provides an intelligent screening and quantitative evaluation method and device for ancient textile diseases. The method comprises the steps of inputting an RGB image and a near infrared image of a panoramic textile to be detected into a trained disease detection and material classification dual-task model based on YOLO, obtaining high-confidence-level boundary box information and low-confidence-level boundary box information, inputting the high-confidence-level boundary box information and the low-confidence-level boundary box information into a material adaptation type pixel level classification model based on DeepLab, obtaining a pixel level segmentation mask and a single-channel-level disease segmentation mask, carrying out boundary correction according to the high-confidence-level boundary box information, the low-confidence-level boundary box information, the disease confidence level, the pixel level segmentation mask, the single-channel-level disease segmentation mask and a disease symbiotic probability matrix, forming a final disease detection segmentation result and a visual annotation image, and generating damage quantization information. The application can obviously improve the accuracy of identifying the disease types.

Inventors

LIU DAWEI
GU YUSHAN
ZHAO RUI
Yuan Suyu

Assignees

中国社会科学院考古研究所

Dates

Publication Date: 20260505
Application Date: 20251208

Claims (10)

1. The intelligent screening and quantitative evaluation method for the ancient textile diseases is characterized by comprising the following steps of: step 1, acquiring an RGB image, a near infrared image and a microscopic image of a panoramic textile to be detected; Step 2, obtaining a trained disease detection and material classification dual-task model based on YOLO and a trained material adaptation type pixel level classification model based on DeepLab; step 3, inputting the RGB image and the near infrared image of the panoramic textile to be detected into the trained disease detection and material classification dual-task model based on the YOLO, so as to obtain high-confidence boundary box information, low-confidence boundary box information, disease confidence, ancient textile disease symbiotic probability matrix and material classification result information; Inputting the high-confidence boundary box information and the low-confidence boundary box information into the DeepLab-based texture adaptive pixel level segmentation model, so as to obtain a pixel level segmentation mask, a single-channel level disease segmentation mask and a texture matching report; Step 5, carrying out boundary correction according to the high confidence boundary box information, the low confidence boundary box information, the disease confidence, the pixel level segmentation mask, the single-channel level disease segmentation mask and the disease symbiotic probability matrix, so as to form a final disease detection segmentation result and a visual labeling image; And 6, generating damage quantification information according to the final disease detection segmentation result, the microscopic image and the material matching report.
2. The intelligent screening and quantitative assessment method for ancient textile diseases according to claim 1, wherein after the step 1 and before the step2, the intelligent screening and quantitative assessment method for ancient textile diseases further comprises: Extracting periodic fiber texture features in an RGB image and a near infrared image through Fourier transformation, constructing a texture mask, carrying out smoothing treatment on textures of a non-disease area, and reserving abnormal textures of the disease area, so as to obtain an RGB image without texture interference, RGB image texture mask information, a near infrared image without texture interference and near infrared image texture mask information; Marking a later repair area by comparing the reflectivity difference of the RGB image with the texture interference, and generating a repair area mask to eliminate the interference of the area on disease detection, thereby obtaining a RGB image with the RGB image exclusive repair area mask, an RGB image repair area coordinate table, a near infrared image with the near infrared image exclusive repair area mask and a near infrared image repair area coordinate table; Respectively carrying out multi-mode characteristic space-time alignment processing on the de-texture interference RGB image containing the RGB image exclusive repair area mask and the de-texture interference near infrared image containing the near infrared image exclusive repair area mask, thereby obtaining a space-time aligned RGB image and a near infrared image after space-time alignment; and 3, inputting the RGB image after space-time alignment and the near infrared image after space-time alignment into the trained disease detection and material classification dual-task model based on the YOLO.
3. The intelligent screening and quantitative assessment method for ancient textile diseases according to claim 2, wherein after the step 3 and before the step 4, the intelligent screening and quantitative assessment method for ancient textile diseases further comprises: Generating high-association high-confidence boundary frame information, high-association high-confidence retention basis tables, high-association low-confidence boundary frame information, high-association low-confidence retention basis tables, low-association high-confidence boundary frame information, low-association high-confidence weakening logs, low-association low-confidence boundary frame information, low-association low-confidence weakening logs, unassociated high-confidence boundary frame information, unassociated high-confidence rejection list, unassociated low-confidence boundary frame information and unassociated low-confidence rejection list according to the high-confidence boundary frame information, low-confidence boundary frame information and material classification result information; And 4, inputting the high-correlation high-confidence boundary box information into the texture adaptive pixel level classification model based on DeepLab by using the high-correlation low-confidence boundary box information.
4. The intelligent screening and quantitative assessment method for ancient textile diseases according to claim 3, wherein after the step 4, before the step 5, the intelligent screening and quantitative assessment method for ancient textile diseases further comprises: Performing protein fiber segmentation mask optimization on the pixel-level segmentation mask and the single-channel-level lesion segmentation mask, so as to obtain a pixel-level segmentation mask after protein optimization and a single-channel-level lesion segmentation mask after protein optimization; Performing cellulose fiber segmentation mask optimization on the pixel-level segmentation mask and the single-channel-level disease segmentation mask so as to obtain a cellulose-optimized pixel-level segmentation mask and a cellulose-optimized single-channel-level disease segmentation mask; Generating a material and disease suitability scoring table according to the protein optimized pixel level segmentation mask, the protein optimized single-channel level disease segmentation mask, the cellulose optimized pixel level segmentation mask and the cellulose optimized single-channel level disease segmentation mask; The step 5 comprises the following steps: And carrying out boundary correction according to the material and disease suitability scoring table, the high-confidence boundary box information, the low-confidence boundary box information, the disease confidence, the pixel-level segmentation mask, the single-channel-level disease segmentation mask and the disease symbiotic probability matrix, so as to form a final disease detection segmentation result and a visual labeling image.
5. The intelligent screening and quantitative evaluation method for ancient textile diseases according to claim 4, wherein the steps of extracting periodic fiber texture features in RGB image and near infrared image by fourier transform, constructing texture masks, smoothing textures of non-disease areas, and retaining abnormal textures of disease areas, thereby obtaining texture mask information of RGB image and RGB image, near infrared image and near infrared image after removing texture interference, comprise: generating a gray RGB image according to the RGB image; preprocessing the RGB image and the near infrared image after the graying, so as to obtain the RGB image and the near infrared image after the preprocessing after the graying after the preprocessing; Respectively constructing an image pyramid for the preprocessed RGB image and the preprocessed near infrared image, so as to obtain an RGB image multi-scale pyramid and a near infrared image multi-scale pyramid; Extracting multi-scale frequency domain texture features of the RGB image multi-scale pyramid and the near infrared image multi-scale pyramid respectively, so as to obtain an RGB image frequency domain matrix set, a near infrared image frequency domain matrix set, an RGB image multi-scale peak coordinate table, a near infrared image multi-scale peak coordinate table, an RGB image multi-scale high frequency energy matrix and a near infrared image multi-scale high frequency energy matrix; Generating a periodic parameter according to the RGB image multi-scale peak coordinate table and the near infrared image multi-scale peak coordinate table; Generating an RGB image weight matrix, a near infrared image weight matrix and weight parameter information according to the RGB image multi-scale high-frequency energy matrix, the near infrared image multi-scale high-frequency energy matrix and the period parameter; and respectively generating an enhanced de-texture RGB image and an enhanced de-texture near infrared image according to the RGB image weight matrix, the near infrared image weight matrix, the preprocessed gray RGB image, the near infrared image and the periodic parameters.
6. The intelligent screening and quantitative evaluation method for ancient textile diseases according to claim 5, wherein the generating of the RGB image weight matrix, the near infrared image weight matrix and the weight parameter information according to the RGB image multi-scale high frequency energy matrix, the near infrared image multi-scale high frequency energy matrix and the period parameter adopts the following formula: when calculating an RGB image, the following formula is adopted: ; Wherein, the The method is characterized in that the method is a smooth weight value at the position of coordinates (x, y) in an RGB image, wherein alpha is a material adaptation coefficient; the actual calculated ancient textile fiber texture period is obtained; Taking the median value of the standard period range for the standard fiber texture period of the corresponding material in the ancient textile material standard database; The tolerance is used for measuring the allowable deviation range of the actual texture period and the standard texture period; High frequency energy values at coordinates (x, y) in the RGB image; taking the median value of a threshold range for the high-frequency energy threshold value of the corresponding disease in the ancient textile disease feature library; the energy attenuation coefficient is used for controlling the influence degree of high-frequency energy on the weight; The texture mask value for an RGB image is used to distinguish between normal texture regions and potentially diseased regions.
7. The intelligent screening and quantitative assessment method for ancient textile diseases according to claim 6, further comprising: Obtaining a material-disease association rule base, wherein the material-disease association rule base comprises material basic information, disease classification and characteristic parameters and material-disease association degree core rule information; The generating high-association high-confidence bounding box information, high-association low-confidence bounding box information, low-association high-confidence weakening log, low-association low-confidence bounding box information, low-association low-confidence weakening log, unassociated high-confidence bounding box information, unassociated high-confidence rejection list, unassociated low-confidence bounding box information, unassociated low-confidence rejection list according to the high-confidence bounding box information, low-association high-confidence bounding box information and material classification result information comprises: Generating a high-correlation boundary box threshold according to the material-disease correlation rule base; classifying each high-confidence-degree boundary box information and each low-confidence-degree boundary box information according to a high-association boundary box threshold value, so as to generate a first high-association candidate boundary box table based on the high-confidence-degree boundary box information; generating a second high-association candidate boundary frame table based on the low-confidence boundary frame information, wherein a part which does not belong to the first high-association candidate boundary frame table in the high-confidence boundary frame information and a part which does not belong to the second high-association candidate boundary frame table in the low-confidence boundary frame information form a non-high-association candidate boundary frame table; Performing multi-mode feature extraction on the first high-association candidate boundary frame table and the second high-association candidate boundary frame table respectively, so as to obtain the first high-association candidate boundary frame table and the second high-association candidate boundary frame table; And carrying out high-correlation boundary frame confidence correction on the first high-correlation candidate boundary frame table and the second high-correlation candidate boundary frame table respectively, so as to obtain high-correlation high-confidence boundary frame information and high-correlation low-confidence boundary frame information.
8. The method for intelligently screening and quantitatively evaluating ancient textile diseases according to claim 7, wherein the performing high-correlation bounding box confidence correction on the first high-correlation bounding box table and the second high-correlation bounding box table respectively to obtain high-correlation high-confidence bounding box information and high-correlation low-confidence bounding box information comprises: calculating confidence information of each first high-association candidate boundary box or each second high-association candidate boundary box respectively; judging whether each first high-association candidate boundary box exceeds a confidence coefficient first classification threshold value, if so, dividing the first high-association candidate boundary boxes exceeding the confidence coefficient first classification threshold value into high-association high-confidence coefficient boundary box information; And judging whether the confidence degree information obtained by each second high-association candidate boundary box exceeds a confidence degree second classification threshold value, and dividing the second high-association candidate boundary boxes exceeding the confidence degree second classification threshold value into high-association low-confidence degree boundary box information.
9. The intelligent screening and quantitative assessment method for ancient textile diseases according to claim 8, wherein the confidence information for calculating each of the first high-association candidate bounding box table or the second high-association candidate bounding box table is calculated by the following formula: ; Is confidence information; initial confidence level output by the disease detection and material classification dual-task model based on YOLO; r is disease-material association degree; the characteristic intensity of the disease area in the RGB image; The characteristic intensity of the corresponding region in the near infrared image; Is a smoothing factor; Is a symbiotic enhancement coefficient; the number of other disease types currently detected; the symbiotic probability of the k-th disease and the current disease; The credibility weight of the k disease is given.
10. The utility model provides an ancient fabrics disease intelligent screening and quantitative evaluation device which characterized in that, ancient fabrics disease intelligent screening and quantitative evaluation device includes: the image acquisition module is used for acquiring RGB image, near infrared image and microscopic image of the panoramic textile to be detected; The model acquisition module is used for acquiring a trained disease detection and material classification dual-task model based on YOLO and a trained material adaptation type pixel level classification model based on DeepLab; The YOLO model processing module is used for inputting an RGB image and a near infrared image of the panoramic textile to be detected into the trained disease detection and material classification dual-task model based on YOLO, so as to obtain high-confidence bounding box information, low-confidence bounding box information, disease confidence, ancient textile disease symbiotic probability matrix and material classification result information; DeepLab a model processing module, wherein the DeepLab model processing module is configured to input the high-confidence bounding box information and the low-confidence bounding box information to the texture adaptive pixel level segmentation model based on DeepLab, so as to obtain a pixel level segmentation mask, a single-channel level lesion segmentation mask and a texture matching report; The detection result acquisition module is used for carrying out boundary correction according to the high-confidence boundary box information, the low-confidence boundary box information, the disease confidence, the pixel-level segmentation mask, the single-channel-level disease segmentation mask and the disease symbiotic probability matrix so as to form a final disease detection segmentation result and a visual labeling image; And the quantization module is used for generating damage quantization information according to the final disease detection segmentation result, the microscopic image and the material matching report.

Description

Intelligent screening and quantitative evaluation method and device for ancient textile diseases Technical Field The invention relates to the technical field of image recognition, in particular to an intelligent screening and quantitative evaluation method and device for ancient textile diseases. Background The ancient textiles are used as important cultural relic carriers for carrying historical cultural information, are mainly made of protein fibers (such as silk and wool) or cellulose fibers (such as hemp and cotton), and are subjected to natural aging, environmental corrosion (such as temperature and humidity fluctuation and microorganism breeding) and human intervention (such as improper repair and storage) for hundreds of years to thousands of years, so that various diseases such as breakage, mildew, fading, worm damage, pollution and the like are easily generated. The diseases not only can damage the physical structure and appearance integrity of the textile, but also can lead to the permanent loss of the historic pattern and color information carried by the textile, so that the accurate detection and quantitative evaluation of the ancient textile diseases are core technical problems to be solved urgently in the field of cultural relic protection, and are also key preconditions for making a scientific repair scheme and prolonging the service life of the cultural relics. Early disease detection relies on visual observation and empirical judgment of cultural relics protection specialists, and is assisted by simple tools such as a magnifying glass, a microscope and the like. The method has a plurality of defects, such as strong subjectivity and poor consistency, and the identification of diseases (such as distinguishing mildew from pollution and judging fracture degree) completely depends on personal experiences of experts, so that different experts often have differences in disease judgment results of the same textile, and lack of unified quantification standards; The method has the advantages of low efficiency and limited coverage range, and is difficult to avoid missing inspection because of the fact that for large-format panoramic textiles (such as ancient brocade and mural fragments), manual area-by-area detection needs days or even weeks, and physical damage risks exist, and part of fragile textiles (such as unearthed silk fragments) cannot bear repeated turning, short-distance observation and other operations, so that fiber breakage or pattern falling can be aggravated by manual detection. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an intelligent screening and quantitative evaluation method for ancient textile diseases, which solves the technical problems in the background art. In order to achieve the above purpose, the present invention provides the following technical solutions: An intelligent screening and quantitative evaluation method for ancient textile diseases comprises the following steps: step 1, acquiring an RGB image, a near infrared image and a microscopic image of a panoramic textile to be detected; Step 2, obtaining a trained disease detection and material classification dual-task model based on YOLO and a trained material adaptation type pixel level classification model based on DeepLab; step 3, inputting the RGB image and the near infrared image of the panoramic textile to be detected into the trained disease detection and material classification dual-task model based on the YOLO, so as to obtain high-confidence boundary box information, low-confidence boundary box information, disease confidence, ancient textile disease symbiotic probability matrix and material classification result information; Inputting the high-confidence boundary box information and the low-confidence boundary box information into the DeepLab-based texture adaptive pixel level segmentation model, so as to obtain a pixel level segmentation mask, a single-channel level disease segmentation mask and a texture matching report; Step 5, carrying out boundary correction according to the high confidence boundary box information, the low confidence boundary box information, the disease confidence, the pixel level segmentation mask, the single-channel level disease segmentation mask and the disease symbiotic probability matrix, so as to form a final disease detection segmentation result and a visual labeling image; And 6, generating damage quantification information according to the final disease detection segmentation result, the microscopic image and the material matching report. Optionally, after the step 1 and before the step 2, the intelligent screening and quantitative evaluation method for ancient textile diseases further comprises: Extracting periodic fiber texture features in an RGB image and a near infrared image through Fourier transformation, constructing a texture mask, carrying out smoothing treatment on textures of a non-disease area, and reserving abnormal te