Lace pattern picture searching system and method

Abstract

The invention discloses a lace pattern picture searching system and method. The system comprises a picture acquisition module, a preprocessing module, a model training module, a feature extraction and storage module, a vector database module and a search processing module. The acquisition module acquires sample images by using various devices, and a quality detection sub-module is arranged in the acquisition module. The preprocessing module is used for carrying out gray level conversion, noise elimination, edge enhancement and self-adaptive cutting on the image, and the multi-spectral texture analysis and real-time deformation compensation sub-module is used for extracting deep texture information and compensating deformation errors. And the search processing module processes the picture uploaded by the user, then performs approximate nearest neighbor search and outputs a result. The method comprises the steps of picture collection, preprocessing, model training, feature extraction and storage and search processing. According to the method, the accuracy and efficiency of lace pattern picture searching are improved through various technical means, the lace pattern pictures can be quickly and accurately searched, accurate recognition can be performed even if deformation exists, and the production order searching speed is increased.

Description

Technical Field

[0001] The present invention relates to the field of textile applications, and in particular to a lace flower shape picture search system and method. Background Art

[0002] Lace is widely used in a variety of fields, including fashion and home decor, and its patterns are diverse and complex. Efficient search for lace pattern images is crucial for lace design, production, sales, and consumption. During production, manufacturers must ensure that lace patterns match customer samples or designs. This requires accurate search and identification of lace patterns to avoid production errors. Furthermore, during quality control, the ability to quickly locate standard patterns that resemble defective products facilitates analysis of the causes of defective products.

[0003] When searching for common lace pattern images, traditional search systems may lack effective feature encoding and database indexing structures when faced with a large amount of lace pattern data. Without efficient feature encoding technology, it is difficult to accurately represent the characteristics of lace patterns, thus affecting search accuracy. Without a suitable database indexing structure, search speed will be very slow, and results similar to the query pattern cannot be quickly located in a large amount of data, which will seriously affect work efficiency in actual production order searches. Moreover, since lace patterns often have complex appearances, relying solely on traditional RGB image information makes it difficult to distinguish patterns that are similar in appearance but different in microstructure. This is because traditional image feature extraction methods cannot deeply explore deep-level features such as the microfiber structure of lace patterns, resulting in easy misjudgment when searching for similar patterns, failing to meet high-precision search requirements and the work requirements of textile applications. To this end, a lace pattern image search system and method are proposed. Summary of the Invention

[0004] The present invention provides the following technical solution: a lace flower image search system, comprising: An image acquisition module, an image preprocessing module, a model training module, a feature extraction and storage module, a vector database, and a search processing module. The image acquisition module includes a scanner, a digital camera, and a three-dimensional imaging device. The image acquisition module is used to acquire lace flower sample images; An image preprocessing module is used to perform grayscale conversion, noise removal, edge enhancement, and adaptive cropping on the captured images. The model training module is based on the ResNet50 residual neural network pre-training model. After removing the fully connected layer, it performs transfer learning to train the model and saves the trained model as a local model file. The image preprocessing module is internally equipped with multispectral texture analysis and real-time deformation compensation submodules. The multispectral texture analysis submodule is used to capture the microscopic fiber structure characteristics of lace patterns through short-wave infrared imaging and combine it with visible light images for multimodal fusion to extract deep texture information that cannot be obtained with traditional RGB images; A feature extraction and association storage module is used to extract features of lace pattern images using a trained model, vectorize the features, and store the vectorized features in a vector database. The feature extraction and storage module uses bio-inspired feature encoding technology, and the vector database adopts a hierarchical index structure. The vector database is used to store feature vectors and associated metadata of lace patterns. The real-time deformation compensation submodule predicts the deformation pattern of fabric under different tensions based on a physical simulation model, and is used to automatically compensate for possible query image deformation errors during the feature extraction stage. The search processing module is used to automatically crop and grayscale the pictures uploaded by users, and at the same time use the model file to extract features, vectorize them, and then transfer them into the vector database for approximate nearest neighbor search, and output the search results from high to low in terms of approximation.

[0005] The present invention provides a lace flower shape picture search method, based on the above-mentioned lace flower shape picture search system, comprising the following steps: S1 image collection steps: The scanner, digital camera, and 3D imaging device in the image acquisition module are used to capture sample images of lace patterns. The scanner can scan according to the size and characteristics of the lace patterns, the digital camera can capture the lace patterns from different angles, and the 3D imaging device uses structured light 3D imaging technology to project a specific pattern onto the lace patterns. The camera captures the deformation of the reflected light pattern to obtain 3D shape information. The captured images are immediately sent to the image quality detection submodule within the image acquisition module, which detects the image clarity, color accuracy, and integrity. S2 image preprocessing steps: The qualified image collected in step S1 is sent to the image preprocessing module for grayscale conversion, noise removal, and edge enhancement. At the same time, the multispectral texture analysis submodule is used to reveal the characteristics of the lace flower shape from different angles. Finally, the real-time deformation compensation submodule automatically compensates for the existing query image deformation error. S3 model training steps: The model training module is based on the ResNet50 residual neural network pre-training model, and transfer learning is performed to train the model after removing the fully connected layer; S4 feature extraction and storage steps: The trained model is used to extract features. A multi-scale feature extraction method is used to extract features from lace flower images at different scales, and the feature vectors are stored in a vector database. S5 search processing steps: After the search processing module receives the picture uploaded by the user, it first automatically crops the picture and performs grayscale processing, uses the previously trained model file to extract features from the processed user picture, and vectorizes the features. The vectorized user picture features are passed into the vector database for approximate nearest neighbor search, and the search results are output from high to low in order of approximation.

[0006] Preferably, the picture acquisition module is internally equipped with an image quality detection submodule, which is used to detect the clarity, color accuracy and integrity of the image immediately after the image is acquired. For images that do not meet the preset quality standards, it prompts re-acquisition and automatic repair processing.

[0007] Preferably, the grayscale conversion process in the image preprocessing module adopts an adaptive grayscale mapping algorithm, and the noise elimination process in the image preprocessing module adopts a filtering algorithm and a noise estimation model based on machine learning.

[0008] Preferably, the edge enhancement in the image preprocessing module adopts an algorithm based on directional derivatives, and the multispectral texture analysis submodule in the image preprocessing module adopts short-wave infrared imaging and long-wave infrared imaging technology.

[0009] Preferably, the real-time deformation compensation submodule in the image preprocessing module is based on the elastic modulus, Poisson's ratio and thickness parameters of the fabric, and a self-learning function is provided inside the real-time deformation compensation submodule.

[0010] Preferably, the grayscale conversion process in step S2 adopts a grayscale conversion algorithm based on local contrast enhancement, and the grayscale conversion algorithm based on local contrast enhancement divides the image into multiple local areas, calculates the contrast of each area separately, and then adjusts the grayscale value according to the regional contrast.

[0011] Preferably, the edge enhancement in step S2 adopts an algorithm based on edge gradient and direction weighting. The algorithm first calculates the gradient size and direction of the edge of the lace flower image, and then performs weighted enhancement according to the importance of the edge in the flower shape and the edge direction, so that the edge enhancement effect is more in line with the characteristics of the lace flower shape.

[0012] Preferably, when a multi-scale feature extraction method is used in the feature extraction and storage step S4, the lace flower pattern is divided according to different structural levels, a smaller scale is used for the flower part of the lace flower pattern, a moderate scale is used for the connecting part between the flowers, and a larger scale is used for the layout of the entire lace flower pattern.

[0013] Preferably, when the feature vectors are stored in the vector database in step S4, the scale information corresponding to each feature vector and the corresponding supplier and copyright holder information are stored synchronously.

[0014] In summary, compared with the prior art, the present invention provides a lace flower image search system and method, which has the following beneficial effects: 1. The present invention uses an image preprocessing module to perform grayscale conversion, noise elimination, edge enhancement, and adaptive cropping on the collected images. Grayscale conversion enhances the image contrast, noise elimination reduces interference factors in the image, edge enhancement highlights the outline of the pattern, and adaptive cropping removes unnecessary background, thereby helping to improve the quality of the pattern image and laying the foundation for subsequent accurate search. In addition, the added multispectral texture analysis submodule captures microscopic fiber structure characteristics through short-wave infrared imaging and multimodally fuses them with visible light images to extract deep texture information that cannot be obtained with traditional RGB images. This enables the system to distinguish lace patterns that are similar in appearance but different in microstructure, greatly improving the accuracy of pattern search. 2. The present invention uses bio-inspired feature encoding technology through a feature extraction and association storage module to effectively extract and vectorize the features of lace pattern images. The vector database uses a hierarchical index structure to store feature vectors and associated metadata, which is conducive to rapid search and matching. Through these technologies, the system can quickly locate the result most similar to the query pattern in a large amount of lace pattern data, achieving fast and accurate search. The added real-time deformation compensation submodule predicts the deformation pattern of the fabric under different tensions based on a physical simulation model, and automatically compensates for the deformation error of the query image during the feature extraction stage. This function is very important for the pattern deformation that may occur in actual production. It ensures that even deformed pattern images can be accurately identified and searched, which improves the speed of production order search. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the system structure of the present invention.

[0016] Figure 2 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] See also Figure 1 The present invention provides a technical solution, a lace flower image search system, comprising: The image acquisition module, image preprocessing module, model training module, feature extraction and storage module, vector database and search processing module are included in the image acquisition module. The image acquisition module includes a scanner, a digital camera and a three-dimensional imaging device. The image acquisition module is used to acquire lace flower sample images. The image acquisition module is internally equipped with an image quality detection submodule. The image quality detection submodule is used to detect the clarity, color accuracy and integrity of the image immediately after acquisition. For images that do not meet the preset quality standards, it prompts re-acquisition and automatic repair processing. The specific workflow of the image quality detection submodule is as follows; Testing starts after image acquisition: In the S1 image acquisition step, after the lace flower sample image is captured using the scanner, digital camera, or 3D imaging device in the image acquisition module, the captured image is immediately sent to the image quality detection submodule within the image acquisition module. At this time, the image quality detection submodule begins to perform multiple tests on the image, mainly including three aspects: clarity, color accuracy, and integrity. Clarity Detection: The image quality detection submodule uses specific algorithms to assess image clarity. It may calculate the sharpness of the edges and details of lace patterns in the image. For example, it measures clarity by detecting the gradient change of the lace pattern's edges. If the gradient change is within a certain range, the clarity is considered to meet the standard. If the gradient change is too small, it may indicate blurry images and substandard clarity. If the image clarity does not meet the preset quality standard, the image quality detection submodule will prompt the user to recapture the image. This prompt may clearly indicate the lack of image clarity and recommend using a more suitable acquisition device or adjusting acquisition parameters. For example, for a digital camera, it may prompt to adjust the focus or stabilize the camera; for a scanner, it may prompt to adjust the scanning resolution. At the same time, the image quality detection submodule will attempt to automatically repair the image. This automatic repair may use an image sharpening algorithm, such as a convolution kernel-based sharpening algorithm, to enhance the contrast of the image edges and improve the image clarity. The repaired image will undergo another clarity check. If the repaired image meets the preset quality standard, it can proceed to the next step. If it still does not meet the standard, it will be prompted to recapture the image. Color accuracy detection: For color accuracy detection, the image quality detection submodule will compare the captured image color with a pre-set standard color reference value. This standard color reference value may be an average value based on the color analysis of a large number of high-quality lace flower images. For example, detect whether the white part of the lace flower shape is within a reasonable white color gamut, whether the flower color is close to the standard color, etc. When it is detected that the color accuracy does not meet the preset quality standards, the image quality detection submodule will prompt to re-capture the image. The prompt content may include suggestions to adjust the color settings of the acquisition device, such as the white balance settings of a digital camera, etc. At the same time, the automatic repair process will be started. Automatic repair may use a color correction algorithm, such as a color correction algorithm based on histogram matching, to adjust the color histogram of the image to a state close to the standard color reference value. The repaired image is tested for color accuracy again. If it meets the standard, it enters the subsequent process, otherwise it will continue to prompt for re-capture; Integrity detection: Integrity detection mainly checks whether there are missing or obscured parts in the image. The image quality detection submodule will analyze the overall structure of the lace flower shape. For example, it will judge the integrity of the image by detecting whether the outline of the lace flower shape is complete, and whether the flowers and lines are broken or missing. If the image integrity does not meet the preset quality standards, the image quality detection submodule will prompt you to re-capture the image and inform you of possible problems. For example, if part of the lace flower shape is obscured, it is recommended to adjust the acquisition angle or remove the obstruction and re-capture. In terms of automatic repair processing, for some small missing parts, if they can be repaired through the image interpolation algorithm, such as the bilinear interpolation algorithm, an attempt will be made to repair them. The repaired image will undergo another integrity test. If it meets the standards, it can enter the subsequent process. Otherwise, it will continue to prompt for re-capture. Process after passing the test: When the image quality detection submodule detects that the image clarity, color accuracy, and integrity meet the preset quality standards, the image will be judged as qualified. The qualified image will enter the S2 image preprocessing step and begin a series of preprocessing operations such as grayscale conversion, noise removal, and edge enhancement; The image preprocessing module is used to perform grayscale conversion, noise removal, edge enhancement, and adaptive cropping on the collected images. The model training module is based on the ResNet50 residual neural network pre-training model. After removing the fully connected layer, it performs transfer learning to train the model and saves the trained model as a local model file. The image preprocessing module is equipped with multispectral texture analysis and real-time deformation compensation submodules. The multispectral texture analysis submodule is used to capture the microscopic fiber structure characteristics of lace patterns through short-wave infrared imaging and combine it with visible light images for multimodal fusion to extract deep texture information that cannot be obtained with traditional RGB images. The grayscale conversion process in the image preprocessing module adopts an adaptive grayscale mapping algorithm, and the noise elimination process in the image preprocessing module adopts a filtering algorithm and a noise estimation model based on machine learning. The specific process of the above method is as follows; Noise Removal Using Filtering Algorithms: Select an appropriate filtering algorithm based on the characteristics and noise type of the lace image. Common filtering algorithms include mean filtering, median filtering, and Gaussian filtering. For lace images, if the noise is primarily salt-and-pepper noise (appearing as random black and white dots in the image), median filtering may be a better choice. If the noise is Gaussian (noise follows a Gaussian distribution), Gaussian filtering is more appropriate. The filtering algorithm parameters are determined based on factors such as image resolution and noise intensity. For example, for median filtering, the filter window size needs to be determined. If the image resolution is high and the noise dots are dispersed, a smaller filter window (such as 3×3) can be selected. If the noise is severe and the image resolution is low, a larger filter window (such as 5×5 or 7×7) can be used. For Gaussian filtering, the size and standard deviation of the Gaussian kernel need to be determined. Generally speaking, a larger Gaussian kernel and standard deviation can better smooth noise, but may cause blurring of image details. This needs to be weighed based on the actual situation. Taking median filtering as an example, for each pixel in the lace flower image, the pixel values within the filter window centered on the pixel are sorted, and then the median value is taken as the new value of the pixel after filtering. For Gaussian filtering, the weight of each pixel in the filter window is calculated according to the Gaussian function, and then the weight is multiplied by the pixel value and summed to obtain the filtered pixel value. By performing such a filtering operation on all pixels in the image, the noise in the image is preliminarily eliminated; Noise removal using a machine learning-based noise estimation model: A large number of noisy lace flower image samples are collected as training data. These image samples need to be annotated with information such as the noise type and intensity. In addition, corresponding noise-free original images are collected as target data. Feature extraction is performed on the image samples, including features such as grayscale value distribution, texture features, and spectral characteristics. These features serve as input to a machine learning model. An appropriate machine learning model is selected, such as a neural network (which can be a simple multilayer perceptron) or a support vector machine (SVM). The model is trained using the extracted image features as input and the annotated information such as the noise type and intensity as output. During training, model parameters, such as the neural network weights and the SVM kernel parameters, are adjusted to minimize the error between the model's predicted noise and the actual annotated noise. Through multiple iterations of training, the model accurately estimates the noise in the image. For the lace flower image to be processed, features are first extracted. The trained machine learning-based noise estimation model is then used to estimate the noise type and intensity in the image. Based on the estimated noise type and intensity, a corresponding noise removal strategy is then applied. For example, if the model estimates that the image contains Gaussian noise with high intensity, a more aggressive Gaussian filter (increasing the Gaussian kernel or standard deviation) can be used; if the image contains salt and pepper noise with low intensity, a median filter with a smaller window can be used. Alternatively, based on the model's predictions, a customized algorithm specifically tailored to that type and intensity of noise can be used for noise removal. Fusion filtering algorithm and machine learning-based noise estimation model: Determine the weight of the filtering algorithm and the machine learning-based noise estimation model in the noise removal process. This weight can be determined based on the image type, prior knowledge of noise, and experimental results. For example, if it is known that the noise type in the image is relatively simple and conforms to a common noise model (such as Gaussian noise), a higher weight (such as 0.7) can be assigned to the filtering algorithm and a lower weight (such as 0.3) can be assigned to the machine learning-based noise estimation model. If the noise type is complex and difficult to predict, the weight of the machine learning-based noise estimation model can be appropriately increased. The edge enhancement in the image preprocessing module uses an algorithm based on directional derivatives, and the multispectral texture analysis submodule in the image preprocessing module uses short-wave infrared imaging and long-wave infrared imaging technology; The real-time deformation compensation submodule in the image preprocessing module is based on the elastic modulus, Poisson's ratio and thickness parameters of the fabric. The real-time deformation compensation submodule is internally equipped with a self-learning function, and the specific process of the self-learning function is as follows; Initialization phase: During system initialization, the real-time deformation compensation submodule sets initial parameters based on the basic properties of the lace fabric, including the elastic modulus, Poisson's ratio, and thickness. These parameters provide the foundation for subsequent deformation prediction and compensation. Simultaneously, a sample dataset is initialized, containing images of known lace patterns and their deformation under different tensions. This sample data can be pre-collected and annotated for training and optimizing the self-learning function. Data acquisition and preprocessing: In the S1 image acquisition step and the S2 image preprocessing step, when the lace flower image is acquired and processed, the real-time deformation compensation submodule collects data related to image deformation. This includes environmental information during image acquisition (such as temperature, humidity, and other factors that may affect fabric deformation), relevant parameters of the acquisition equipment (such as the intensity of the projected light of the three-dimensional imaging device, etc.), and the characteristic information of the image itself (such as the texture and structure of the lace flower). The collected data is preprocessed, and the values of the environmental information, equipment parameters, etc. are standardized so that they are within the appropriate numerical range to facilitate subsequent calculations and analysis. For image feature information, key feature vectors are extracted. For example, a bio-inspired feature encoding technology similar to the feature extraction and storage module is used to convert the image features of the lace flower into a computable vector form; Deformation Prediction and Compensation: During the feature extraction phase (e.g., S4 Feature Extraction and Storage), while processing an image, the real-time deformation compensation submodule predicts the deformation pattern of the fabric under different tensions based on an established physical simulation model. This physical simulation model is constructed based on the fabric's elastic modulus, Poisson's ratio, and thickness parameters. By inputting environmental information and equipment parameters corresponding to the current image, it calculates the theoretically possible deformation scenarios and automatically compensates the query image based on the predicted deformation patterns. For example, if it is predicted that the lace pattern in the image may be stretched in a certain direction due to tension, the feature extraction process adjusts the feature vector calculation method or performs corresponding geometric transformations on the image to offset this deformation, ensuring that the extracted features accurately reflect the original lace pattern. After deformation prediction and compensation are completed, the compensated image features are compared with the actual results. The actual results can be inferred from the accurate matching results obtained in the subsequent search process (e.g., S5 Search Processing). If it is found that the compensated image features still cannot match the correct lace flower shape well (for example, the search result has a low approximation), the collected data (including pre-processed environmental information, equipment parameters, image features, etc.) and the relevant parameters in the compensation process (such as the predicted deformation variable, the actual compensation method used, etc.) are used as feedback data, and the physical simulation model is adjusted using these feedback data. For example, if it is found that the actual deformation variable of a certain fabric in a specific environment deviates from the deformation variable predicted by the model, then the relevant coefficients in the physical simulation model, such as the correction coefficient of the elastic modulus, are adjusted so that the model can predict the deformation more accurately. At the same time, the algorithm and parameters of the feature compensation are optimized based on the feedback data. For example, if it is found that a certain type of lace flower shape requires a different compensation strategy under a specific collection device to obtain better results, then the relevant parameters in the feature compensation algorithm, such as the compensation weight, the transformation ratio, etc., are adjusted; Continuous learning and optimization: The real-time deformation compensation submodule regularly updates the sample dataset (for example, after a certain number of image acquisition and processing cycles). Newly acquired images and their related data (including cases of successful and unsuccessful compensation) are added to the sample dataset to increase the diversity and representativeness of the samples. The updated sample dataset is then used to retrain the physical simulation model and optimize the feature compensation algorithm. In this way, the real-time deformation compensation submodule can continuously adapt to different types of lace flower shapes, different acquisition environments, and equipment, improving the accuracy of deformation prediction and compensation. During daily image acquisition and processing, the real-time deformation compensation submodule continuously monitors its own performance. For example, by statistically analyzing the distribution of the approximation of search results, if it is found that the approximation of search results is generally low over a period of time, it may indicate that there is a problem with inaccurate deformation compensation. When performance degradation is detected, more frequent data acquisition and analysis processes are automatically triggered to accelerate the speed of self-learning and more quickly optimize the physical simulation model and feature compensation algorithm to restore or improve the accuracy of deformation compensation. The feature extraction and association storage module is used to extract features of lace patterns using the trained model, vectorize the features, and store the vectorized features in a vector database. The feature extraction and storage module uses bio-inspired feature encoding technology, and the vector database adopts a hierarchical index structure. The vector database is used to store the feature vectors of lace patterns and associated metadata. The real-time deformation compensation submodule predicts the deformation pattern of the fabric under different tensions based on a physical simulation model, and is used to automatically compensate for possible query image deformation errors during the feature extraction stage. The search processing module is used to automatically crop and grayscale the images uploaded by users. At the same time, it uses the model file to extract features, vectorizes them, and then transfers them to the vector database for approximate nearest neighbor search, and outputs the search results in descending order of approximation.

[0019] A lace flower shape image search method, based on the above-mentioned lace flower shape image search system, includes the following steps: S1 image collection steps: The scanner, digital camera, and 3D imaging device in the image acquisition module are used to capture sample images of lace patterns. The scanner can scan according to the size and characteristics of the lace patterns, the digital camera can capture the lace patterns from different angles, and the 3D imaging device uses structured light 3D imaging technology to project a specific pattern onto the lace patterns. The camera captures the deformation of the reflected light pattern to obtain 3D shape information. The captured images are immediately sent to the image quality detection submodule within the image acquisition module, which tests the image clarity, color accuracy, and integrity. S2 image preprocessing steps: The qualified image collected in step S1 enters the image preprocessing module for grayscale conversion, noise elimination and edge enhancement. At the same time, the multispectral texture analysis submodule reveals the characteristics of the lace flower shape from different angles. Finally, the real-time deformation compensation submodule automatically compensates for the existing query image deformation error. The grayscale conversion process adopts a grayscale conversion algorithm based on local contrast enhancement. The grayscale conversion algorithm based on local contrast enhancement divides the image into multiple local areas, calculates the contrast of each area respectively, and then adjusts the grayscale value according to the regional contrast. The edge enhancement adopts an algorithm based on edge gradient and direction weighting. The edge gradient and direction weighting algorithm first calculates the gradient size and direction of the edge of the lace flower image, and then performs weighted enhancement according to the importance of the edge in the flower shape and the edge direction, so that the edge enhancement effect is more consistent with the characteristics of the lace flower shape; S3 model training steps: The model training module is based on the ResNet50 residual neural network pre-training model, and transfer learning is performed to train the model after removing the fully connected layer; S4 feature extraction and storage steps: The trained model is used for feature extraction. A multi-scale feature extraction method is used to extract features from the lace flower image at different scales, and the feature vectors are stored in a vector database. When the multi-scale feature extraction method is used in the feature extraction and storage steps, the lace flower image is divided according to different structural levels. A smaller scale is used for the flower portion of the lace flower image, a moderate scale is used for the connecting portion between the flowers, and a larger scale is used for the layout of the entire lace flower image. When the feature vectors are stored in the vector database, scale information corresponding to each feature vector and corresponding supplier and copyright holder information are simultaneously stored. S5 search processing steps: After the search processing module receives the picture uploaded by the user, it first automatically crops the picture and performs grayscale processing, uses the previously trained model file to extract features from the processed user picture, and vectorizes the features. The vectorized user picture features are passed into the vector database for approximate nearest neighbor search, and the search results are output from high to low in order of approximation.

[0020] This solution uses an image preprocessing module to perform grayscale conversion, noise elimination, edge enhancement, and adaptive cropping on the collected images. Grayscale conversion enhances the image contrast, while noise elimination reduces interference factors in the image, edge enhancement highlights the outline of the pattern, and adaptive cropping removes unnecessary background, which helps to improve the quality of the pattern image and lays the foundation for subsequent accurate search. In addition, the added multispectral texture analysis submodule captures microscopic fiber structure characteristics through short-wave infrared imaging and multimodally fuses them with visible light images to extract deep texture information that cannot be obtained with traditional RGB images. This enables the system to distinguish lace patterns that are similar in appearance but different in microstructure, greatly improving the accuracy of pattern search.

[0021] This solution utilizes bio-inspired feature encoding technology through a feature extraction and associative storage module, effectively extracting and vectorizing the features of lace pattern images. The vector database uses a hierarchical index structure to store feature vectors and associated metadata, which facilitates rapid search and matching. Through these technologies, the system can quickly locate the result most similar to the query pattern within a large amount of lace pattern data, enabling fast and accurate searches. Furthermore, the added real-time deformation compensation submodule predicts the deformation pattern of fabrics under different tensions based on a physical simulation model, automatically compensating for deformation errors in the query image during the feature extraction phase. This function is crucial for the potential pattern deformation that may occur in actual production, ensuring that even deformed pattern images can be accurately identified and searched, speeding up production order searches.

[0022] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0023] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A lace flower image search system, characterized in that: include: An image acquisition module, an image preprocessing module, a model training module, a feature extraction and storage module, a vector database, and a search processing module. The image acquisition module includes a scanner, a digital camera, and a three-dimensional imaging device. The image acquisition module is used to acquire lace flower sample images; An image preprocessing module is used to perform grayscale conversion, noise removal, edge enhancement, and adaptive cropping on the captured images. The model training module is based on the ResNet50 residual neural network pre-training model. After removing the fully connected layer, it performs transfer learning to train the model and saves the trained model as a local model file. The image preprocessing module is internally equipped with multispectral texture analysis and real-time deformation compensation submodules. The multispectral texture analysis submodule is used to capture the microscopic fiber structure characteristics of lace patterns through short-wave infrared imaging and combine it with visible light images for multimodal fusion to extract deep texture information that cannot be obtained with traditional RGB images; A feature extraction and association storage module is used to extract features of lace pattern images using a trained model, vectorize the features, and store the vectorized features in a vector database. The feature extraction and storage module uses bio-inspired feature encoding technology, and the vector database adopts a hierarchical index structure. The vector database is used to store feature vectors and associated metadata of lace patterns. The real-time deformation compensation submodule predicts the deformation pattern of fabric under different tensions based on a physical simulation model, and is used to automatically compensate for possible query image deformation errors during the feature extraction stage. The search processing module is used to automatically crop and grayscale the pictures uploaded by users, extract features using the model file and vectorize them, and then transfer them into the vector database for approximate nearest neighbor search, and output the search results in descending order of approximation.

2. The lace flower shape image search system according to claim 1, characterized in that: The image acquisition module is internally equipped with an image quality detection submodule, which is used to detect the clarity, color accuracy and integrity of the image immediately after the image is acquired. For images that do not meet the preset quality standards, it prompts re-acquisition and automatic repair processing.

3. The lace flower pattern image search system according to claim 1, characterized in that: The grayscale conversion process in the image preprocessing module adopts an adaptive grayscale mapping algorithm, and the noise elimination process in the image preprocessing module adopts a filtering algorithm and a noise estimation model based on machine learning.

4. The lace flower pattern image search system according to claim 1, characterized in that: The edge enhancement in the image preprocessing module adopts an algorithm based on directional derivatives, and the multispectral texture analysis submodule in the image preprocessing module adopts short-wave infrared imaging and long-wave infrared imaging technology.

5. The lace flower pattern image search system according to claim 1, characterized in that: The real-time deformation compensation submodule in the image preprocessing module is based on the elastic modulus, Poisson's ratio and thickness parameters of the fabric, and a self-learning function is provided inside the real-time deformation compensation submodule.

6. A lace flower shape image search method, based on the lace flower shape image search system according to any one of claims 1 to 5, comprising the following steps: S1 image collection steps: The scanner, digital camera, and 3D imaging device in the image acquisition module are used to capture sample images of lace patterns. The scanner can scan according to the size and characteristics of the lace patterns, the digital camera can capture the lace patterns from different angles, and the 3D imaging device uses structured light 3D imaging technology to project a specific pattern onto the lace patterns. The camera captures the deformation of the reflected light pattern to obtain 3D shape information. The captured images are immediately sent to the image quality detection submodule within the image acquisition module, which detects the image clarity, color accuracy, and integrity. S2 image preprocessing steps: The qualified image collected in step S1 is sent to the image preprocessing module for grayscale conversion, noise removal, and edge enhancement. At the same time, the multispectral texture analysis submodule is used to reveal the characteristics of the lace flower shape from different angles. Finally, the real-time deformation compensation submodule automatically compensates for the existing query image deformation error. S3 model training steps: The model training module is based on the ResNet50 residual neural network pre-training model, and transfer learning is performed to train the model after removing the fully connected layer; S4 feature extraction and storage steps: The trained model is used to extract features. A multi-scale feature extraction method is used to extract features from lace flower images at different scales, and the feature vectors are stored in a vector database. S5 search processing steps: After the search processing module receives the picture uploaded by the user, it first automatically crops the picture and performs grayscale processing, uses the previously trained model file to extract features from the processed user picture, and vectorizes the features. The vectorized user picture features are passed into the vector database for approximate nearest neighbor search, and the search results are output from high to low in order of approximation.

7. The lace flower image search method according to claim 1, characterized in that: The grayscale conversion process in step S2 adopts a grayscale conversion algorithm based on local contrast enhancement. The grayscale conversion algorithm based on local contrast enhancement divides the image into multiple local areas, calculates the contrast of each area respectively, and then adjusts the grayscale value according to the regional contrast.

8. The lace flower image search method according to claim 1, characterized in that: The edge enhancement in step S2 adopts an algorithm based on edge gradient and direction weighting. The algorithm first calculates the gradient size and direction of the edge of the lace flower image, and then performs weighted enhancement based on the importance of the edge in the flower shape and the edge direction, so that the edge enhancement effect is more consistent with the characteristics of the lace flower shape.

9. The lace flower image search method according to claim 1, characterized in that: When a multi-scale feature extraction method is used in the feature extraction and storage step S4, the lace pattern is divided according to different structural levels. A smaller scale is used for the flower part of the lace pattern, a moderate scale is used for the connecting part between the flowers, and a larger scale is used for the layout of the entire lace pattern.

10. The lace flower image search method according to claim 1, characterized in that: In step S4, when the feature vectors are stored in the vector database, the scale information corresponding to each feature vector and the corresponding supplier and copyright holder information are stored simultaneously.

Images