Hydropower industry metal or concrete crack identification method based on self-supervised enhancement characteristics

Abstract

The invention provides a hydropower industry metal or concrete crack identification method based on self-supervised enhanced features, and the method achieves a complete data retrieval mode through the technologies of large model pre-training, self-supervised learning and the like, can screen out effective data from mass data, and improves the recognition efficiency based on a large amount of scene data. According to the method, the pre-training model of the hydroelectric scene is trained, the pre-training model has super-strong perception ability for the image data of the hydroelectric scene, and based on the self-supervision pre-training model, the hydroelectric industry metal and concrete crack identification algorithm based on the self-supervision enhanced characteristics is developed, and the accuracy rate and the generalization ability are high.

Description

Technical Field

[0001] The invention relates to the technical field of metal and concrete crack recognition in the hydropower industry, and mainly to a metal or concrete crack recognition method in the hydropower industry based on self-supervision enhancement features. Background Art

[0002] In the hydropower industry, the safety of metal and concrete structures is crucial to the operation of the project. However, these structures may develop damage such as cracks during long-term use, which can seriously affect their function and life. Traditional crack detection methods usually rely on manual visual inspection, which is not only time-consuming and labor-intensive, but also easily affected by the inspector's experience and subjective judgment, resulting in inaccurate and inconsistent detection results.

[0003] At present, there are also adopted: Metal and concrete crack detection technology based on traditional image processing methods. It usually involves using optical instruments or camera equipment to capture images of cracks, and then analyzing them through manual or automated software. The details are as follows: Technical solution: Image acquisition: Use a high-resolution camera or scanning equipment to obtain images of cracks. Image processing: Pre-process the acquired images through image processing software, such as denoising, contrast enhancement, etc. Feature extraction: Extract crack features such as length, width, shape, etc. from the processed image. Crack identification: Based on the extracted features, determine whether there are cracks in the image through manual recognition or automatic classification algorithms. However, metal and concrete crack detection technology based on traditional image processing methods has the following defects: (1) Subjectivity: Image-based crack detection often requires human intervention, and human subjective judgment may affect the accuracy of the detection results.

[0004] (2) Environmental interference: Factors such as lighting, shadows, and reflections may interfere with image acquisition and processing, leading to errors in crack identification.

[0005] (3) Crack morphology diversity: Cracks vary in shape, direction, and width. Existing image processing technology may be difficult to adapt to all types of cracks.

[0006] (4) Poor real-time performance: Traditional image-based processing methods are usually slow and not suitable for real-time monitoring.

[0007] (5) Low degree of automation: Although there is automated software, it still relies on manual operation overall, and the degree of automation is not high.

[0008] Also adopted: Metal and concrete crack identification solution based on computer vision small model Technical route: (1) Data collection and preprocessing: Use high-resolution cameras or drone-mounted cameras to collect images. Apply image preprocessing techniques, such as denoising, contrast enhancement, and image cropping, to improve the accuracy and efficiency of crack recognition.

[0009] (2) Feature extraction: Use edge detection algorithms (such as the Canny operator) to extract edge features of cracks. Use morphological operations (such as corrosion and expansion) to highlight the crack area.

[0010] (3) Model selection and optimization: Select lightweight deep learning models suitable for mobile devices or embedded systems, such as MobileNet, SqueezeNet, etc. Optimize the model structure, such as increasing the depth or width, or use model pruning, quantization and other techniques to improve its ability to capture crack characteristics.

[0011] (4) Data enhancement and transfer learning: Apply data augmentation techniques, such as rotation, scaling, flipping, etc., to expand the training data set and improve the generalization ability of the model. Use transfer learning to transfer the parameters of the pre-trained large model to the crack identification task to improve the performance of the small model.

[0012] (5) Model training and evaluation: Train small models on labeled data sets and use methods such as cross-validation to evaluate model performance. Apply model fusion techniques, such as ensemble learning, to further improve recognition accuracy and anti-interference capabilities.

[0013] However, the technical route of metal and concrete crack recognition based on small computer vision models has the following defects: (1) Poor recognition accuracy: Due to the low complexity of the small model, it may be difficult to capture the subtle features of the crack, especially when the crack width is small or the image quality is poor, the recognition accuracy may decrease.

[0014] (2) Weak anti-interference ability: Small models may be more sensitive to interference factors such as noise and light changes, which may affect the accurate identification of cracks, especially in complex environments.

[0015] (3) Poor generalization ability: Small models may perform poorly on data outside the training dataset and have limited generalization ability, which may limit their application in different scenarios.

[0016] In summary, traditional crack recognition methods are mainly based on small models, which usually only perform well on specific data sets and perform poorly on new or different data sets. Small models have limited learning capabilities and it is difficult to learn complex features and laws from limited data. In addition, small models usually require a large amount of labeled data for training, and obtaining a large amount of high-quality labeled data often requires a lot of manpower and time. Summary of the invention

[0017] In view of the above problems, the present invention provides a method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features. This method is based on a large model and has high accuracy and robustness, and can achieve timely detection and early warning of early micro cracks. In addition, with the help of large-scale data sets and powerful computing power, this method also helps to continuously improve the performance of crack identification and adapt to the high requirements of the hydropower industry for safety and durability. Therefore, this technology has broad application prospects in the hydropower industry.

[0018] In order to realize the above technical features, the purpose of the present invention is achieved as follows: a method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features, comprising: Step S1: Data screening: The scheme using image retrieval technology takes the similarity measurement between images as the basis, and finds images similar to the query image by comparing the features of the images; Step S2: Self-supervised learning: The pre-training of the large model for metal and concrete crack recognition uses natural image data and image reconstruction to divide the image into non-overlapping patches. When inputting the model, a certain proportion of patches are hidden, so that the model can reconstruct the hidden patches based on the input patches and has the ability to extract common features of the image. Step S3: Large model compression: Based on application-level considerations, on the basis of the pre-trained large model of metal and concrete cracks, knowledge distillation and fine-tuning techniques are used to empower the knowledge of the pre-trained large model for downstream tasks to develop metal and concrete recognition models.

[0019] The image retrieval technology in step S1 includes image preprocessing, feature extraction, feature representation, similarity measurement, index construction and query processing.

[0020] The image preprocessing stage includes performing denoising, scaling, and cropping operations on the image.

[0021] The feature extraction is to extract representative features from the image, specifically including color, texture, and shape.

[0022] The feature representation is to express the extracted features in the form of vectors or matrices in order to perform similarity measurement.

[0023] The similarity metric is used to measure the similarity between images, including Euclidean distance and cosine similarity; based on the similarity metric, images are sorted to achieve image retrieval.

[0024] The index construction is to associate image features with images in the image library to form a feature index table, so that similar images can be found directly through the feature index table during query.

[0025] Step S3 also includes pre-training the large model on massive general data, and then transferring the perception capability of the large model to the small model at the application layer through knowledge distillation and fine-tuning technology, so that the small model not only has the perception capability equivalent to the large model, but also facilitates the deployment of the model in actual industrial scenarios.

[0026] The present invention has the following beneficial effects: 1. The present invention adopts a large model, which has stronger learning and generalization capabilities than a small model. Large models usually contain hundreds of millions or even hundreds of billions of parameters and can learn complex features and rules from large-scale data. This enables large models to perform well on different data sets and have better generalization capabilities. In addition, large models can also apply knowledge learned in one task to other related tasks through technologies such as transfer learning, thereby reducing the need for labeled data.

[0027] 2. The metal and concrete crack identification method for the hydropower industry based on the large model of the present invention has high accuracy and robustness, and can achieve timely detection and early warning of early micro cracks. In addition, with the help of large-scale data sets and powerful computing power, this method also helps to continuously improve the performance of crack identification and adapt to the high requirements of the hydropower industry for safety and durability. Therefore, this technology has broad application prospects in the hydropower industry.

[0028] 3. The present invention realizes a complete data retrieval mode through large model pre-training technology, and can filter out valid data from massive data. By self-supervising the classification of data scenes, the filtered data scenes are rich and have low repetition rate.

[0029] 4. The present invention is based on self-supervised learning and a large amount of scene data to train a pre-trained model based on the hydropower scene. The pre-trained model has a super perception ability for the image data of the hydropower scene.

[0030] 5. Based on the self-supervised pre-training model, the present invention develops a metal and concrete crack recognition algorithm based on large model technology, which has high accuracy and strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] Figure 1 It is a flow chart of the method of the present invention.

[0033] Figure 2 It is a data screening flow chart of the present invention.

[0034] Figure 3 This is a metal crack identification effect diagram of the present invention.

[0035] Figure 4 This is a concrete crack identification effect diagram of the present invention. DETAILED DESCRIPTION

[0036] The present invention is described in detail below with reference to specific examples.

[0037] Embodiment 1: See also Figure 1-4 , supervised methods based on small models rely on a large amount of labeled data to train the model, which is not only costly but also time-consuming. Usually, metal and concrete crack defect data are difficult to obtain, and it is extremely difficult to obtain a sufficient amount of labeled data, which greatly limits the application scope and effectiveness of supervised methods. In addition, the large number of industrial scenarios and the variety of crack defect forms also pose great challenges to the scene adaptation and model generalization capabilities of small models. Compared with supervised small models, large model methods mainly solve the problems of recognition accuracy and generalization ability.

[0038] (1) Data screening: A large model requires not only large computing power, but also a large amount of scenario data for verification, so we need a set of complete solution progress and effective data selection (2) Self-supervised learning: In large-model object detection, the primary issue to be addressed is how to effectively use large unlabeled datasets to assist model training in order to improve the performance and generalization of the model. This includes designing appropriate large-model learning algorithms and integrating supervised and self-supervised learning strategies to maximize the value of information in unlabeled data.

[0039] (3) Large model compression: A typical feature of large models is that they have a large number of model parameters. Large models require large computing power for both training and reasoning, so it is usually difficult to deploy large models directly to user end-point testing. How to compress large models so that they can be directly applied to end-point testing is an important direction for the development of large model technology.

[0040] This solution combines the hydropower scenario, and the specific technical planning steps are as follows: Step S1: Data screening: The scheme using image retrieval technology takes the similarity measurement between images as the basis, and finds images similar to the query image by comparing the features of the images. Image retrieval technology includes image preprocessing, feature extraction, feature representation, similarity measurement, index construction and query processing. For example, Figure 2 As shown: In the image preprocessing stage, the main operations are denoising, scaling, cropping, etc. to improve the image quality and facilitate subsequent processing. The next step is feature extraction, which refers to extracting representative features from the image, such as color, texture, shape, etc. These features are the key to image retrieval and directly affect the quality of the retrieval effect. Feature representation is to express the extracted features in a certain form in order to measure similarity. Common representation methods include vectors and matrices. Similarity measurement is a method to measure the similarity between images, mainly Euclidean distance, cosine similarity, etc. According to the similarity measurement, images can be sorted to achieve image retrieval. In order to improve the retrieval speed, it is usually necessary to build an index to associate image features with images in the image library to form a feature index table. In this way, when querying, similar images can be found directly through the feature index table.

[0041] Step S2: Self-supervised learning: The pre-training of the large model for metal and concrete crack recognition uses massive natural image data and image reconstruction to divide the image into non-overlapping patches. A certain proportion of patches are hidden when inputting the model, so that the model can reconstruct the hidden patches based on the input patches and has the ability to extract common features of the image.

[0042] Step S3: Large model compression: Based on application considerations, on the basis of the pre-trained large model of metal and concrete cracks, the knowledge of the pre-trained large model is empowered to downstream tasks using technologies such as knowledge distillation and fine-tuning to develop metal and concrete recognition models. The large model is pre-trained on massive general data, and then the perception ability of the large model is transferred to the small model at the application layer through technologies such as knowledge distillation and fine-tuning, so that the small model not only has the perception ability equivalent to the large model, but also facilitates the deployment of the model in actual industrial scenarios.

[0043] After the above steps, combined with the actual hydropower scene material pictures, the metal / concrete crack recognition effect example is as follows Figure 3 , 4 described.

Claims

1. A method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features, characterized in that: include: Step S1: Data screening: The scheme using image retrieval technology takes the similarity measurement between images as the basis, and finds images similar to the query image by comparing the features of the images; Step S2: Self-supervised learning: The pre-training of the large model for metal and concrete crack recognition uses natural image data and image reconstruction to divide the image into non-overlapping patches. When inputting the model, a certain proportion of patches are hidden, so that the model can reconstruct the hidden patches based on the input patches and has the ability to extract common features of the image. Step S3: Large model compression: Based on application-level considerations, on the basis of the pre-trained large model of metal and concrete cracks, knowledge distillation and fine-tuning techniques are used to empower the knowledge of the pre-trained large model for downstream tasks to develop metal and concrete recognition models.

2. According to claim 1, the method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features is characterized in that: The image retrieval technology in step S1 includes image preprocessing, feature extraction, feature representation, similarity measurement, index construction and query processing.

3. According to claim 2, the method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features is characterized in that: The image preprocessing stage includes performing denoising, scaling, and cropping operations on the image.

4. According to claim 3, the method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features is characterized in that: The feature extraction is to extract representative features from the image, specifically including color, texture, and shape.

5. According to claim 4, the method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features is characterized in that: The feature representation is to express the extracted features in the form of vectors or matrices in order to perform similarity measurement.

6. The method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features according to claim 5 is characterized in that: The similarity metric is used to measure the similarity between images, including Euclidean distance and cosine similarity; based on the similarity metric, images are sorted to achieve image retrieval.

7. The method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features according to claim 6 is characterized in that: The index construction is to associate image features with images in the image library to form a feature index table, so that when querying, similar images can be found directly through the feature index table.

8. The method for identifying metal or concrete cracks in the hydropower industry based on self-supervised enhanced features according to claim 6 is characterized in that: Step S3 also includes pre-training the large model on massive general data, and then transferring the perception capability of the large model to the small model at the application layer through knowledge distillation and fine-tuning technology, so that the small model not only has the perception capability equivalent to the large model, but also facilitates the deployment of the model in actual industrial scenarios.

Images