DETR-based concrete pouring vibration apparent quality evaluation method and system

Abstract

The invention discloses a DETR-based concrete pouring vibration apparent quality evaluation method and system. The method comprises the following steps: acquiring a concrete surface image; performing semantic segmentation and defect detection on images in the data set by adopting a concrete pouring vibration appearance detection model; and performing quantitative scoring on the concrete pouring vibration apparent quality according to a defect detection result. By fusing a deep learning technology and a visual detection means, high-precision automatic identification and quality evaluation of concrete surface defects (such as cracks, honeycombs and holes) are realized, the problems that traditional concrete vibration apparent quality evaluation depends on artificial experience, the efficiency is low and the subjectivity is high can be solved, and the quality evaluation accuracy is improved. The method is suitable for visual defect detection and quality grading based on deep learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of construction engineering quality detection, and more specifically, relates to a method and system for evaluating the apparent quality of concrete pouring and vibration based on DETR. Background Art

[0002] Concrete pouring and vibrating construction is a critical step affecting the safety and durability of structures. Its surface quality directly reflects the standardization of construction techniques and the stability of material properties. Traditional quality assessment methods rely primarily on manual visual inspection or traditional rule-based image processing techniques, which suffer from low efficiency, strong subjectivity, and difficulty in quantifying complex defects. In recent years, deep learning methods based on convolutional neural networks (CNNs) have been introduced into this field. However, these methods rely on anchor box design, are not adaptable enough for detecting multi-scale defects and small objects, and exhibit high model complexity. The Detection Transformer (DETR), an emerging end-to-end object detection framework, effectively overcomes the limitations of traditional methods by globally modeling image features through a self-attention mechanism and generating prediction boxes using parallel decoding. However, existing research has not yet applied DETR to the field of concrete surface quality assessment, particularly for the dynamic and variable surface defect detection scenarios encountered during vibrating construction. Therefore, developing an efficient and robust DETR-based surface quality assessment method is of great significance. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a method and system for evaluating the apparent quality of concrete pouring and vibration based on DETR. By integrating deep learning technology with visual inspection means, high-precision automatic recognition and quality assessment of concrete surface defects (such as cracks, honeycombs, holes, etc.) are achieved.

[0004] The purpose of the present invention can be achieved through the following technical solutions.

[0005] A method for evaluating the apparent quality of concrete pouring and vibration based on DETR, comprising the following steps:

[0006] S1: collect concrete surface images;

[0007] S2: Use the concrete pouring and vibrating surface detection model to perform semantic segmentation and defect detection on the images in the dataset;

[0008] S3: Quantitatively score the apparent quality of concrete pouring and vibration based on the defect detection results.

[0009] Furthermore, in step S1, an industrial camera is used in conjunction with a ring light to capture an image of the concrete surface.

[0010] Furthermore, the concrete pouring vibration appearance detection model described in step S2 adopts a DETR model, and the model structure includes an image input layer, an image preprocessing layer, a CNN backbone network layer, an encoder, a Transformer decoder, a prediction head, a post-processing layer, and an output layer arranged in sequence.

[0011] Furthermore, the image input layer is used to input images in the data set;

[0012] The image preprocessing layer is used to perform normalization and data enhancement on the input image;

[0013] The CNN backbone network layer uses a ResNet network or an EfficientNet network to extract multi-scale features from the preprocessed image, including cracks, honeycombs, and holes, output feature maps, and provide semantic information;

[0014] The encoder includes a position encoding layer and a Transformer encoder; the position encoding layer is used to perform position encoding on the feature map to obtain global dependencies; the Transformer encoder processes the position-encoded sequence data through a multi-head self-attention mechanism and a feedforward network, and outputs global semantics and position information to the Transformer decoder;

[0015] The Transformer decoder performs object queries on the Transformer encoder output features, uses a cross-attention mechanism to interact the object queries with the encoder output features, locates the location of concrete pouring and vibration defects, and uses a self-attention mechanism to interact the object queries directly with each other, learns and distinguishes different concrete pouring and vibration defect categories, and outputs the concrete pouring and vibration defect categories and locations in the form of feature sequences;

[0016] The prediction head is used to classify and regress the feature sequence output by the Transforme decoder, and output the category probability and location coordinates of the concrete pouring and vibration defects;

[0017] The post-processing layer is used to compare the set confidence threshold with the category probability of concrete pouring and vibration defects, and filter out the prediction head output results below the confidence threshold;

[0018] The output layer is used to output the post-processed prediction results, including the category, location, and confidence level of the concrete pouring and vibration defects.

[0019] Furthermore, the training data set acquisition process of the concrete pouring and vibration surface detection model is as follows: an industrial camera is used in conjunction with a ring fill light to collect concrete surface images, which are processed and annotated according to the annotation standard format to construct a data set containing multi-scale concrete surface images;

[0020] Annotation specifications: The collected concrete surface images follow the COCO format for defect annotation, and the annotation categories include cracks, honeycombs, and holes. For cracks on the concrete surface, the detection indicators include: length, width, length error, width error, processing speed, and anti-interference ability; for honeycombs on the concrete surface, the detection indicators include: minimum hole diameter and depth estimation accuracy; for holes on the concrete surface, the detection indicators include: shape adaptability, depth classification, minimum area detection, edge accuracy, and multi-scale detection.

[0021] Furthermore, the image preprocessing layer performs data enhancement on the input image, including random cropping, Gaussian noise injection and MixUp enhancement.

[0022] Furthermore, a feature pyramid network is introduced at the front end of the encoder to fuse convolutional features at different levels and realize multi-scale feature fusion.

[0023] Furthermore, during the training process of the concrete pouring vibration surface detection model, the Transformer decoder performs dynamic query initialization before performing object query, queries the training image through engineering prior knowledge, and finds high-frequency defect areas.

[0024] Furthermore, the loss function used in the training process of the concrete pouring vibration appearance detection model is a hybrid loss function, which combines classification cross entropy loss and GIoU positioning loss.

[0025] The purpose of the present invention can also be achieved through the following technical solutions.

[0026] A concrete pouring and vibrating surface quality evaluation system based on DETR, including:

[0027] Image acquisition module, collecting concrete surface images;

[0028] The defect detection module uses a concrete pouring and vibration surface detection model to perform semantic segmentation and defect detection on images in the dataset;

[0029] The quality evaluation module quantitatively scores the apparent quality of concrete pouring and vibration based on the defect detection results.

[0030] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0031] (1) This invention constructs a dataset containing multi-scale concrete surface images and optimizes model training strategies, providing technical support for real-time monitoring and intelligent management of concrete construction quality. Furthermore, the collected concrete surface images must be annotated with defects in the COCO format before they can be used as training datasets.

[0032] (2) The present invention introduces a feature pyramid network (FPN) at the front end of the encoder to fuse convolutional features at different levels, realize multi-scale feature fusion, and improve the detection sensitivity of small defects.

[0033] (3) During the model training process, the present invention performs dynamic query initialization before the Transformer decoder performs object query. The training image is queried based on engineering prior knowledge. In the initialization stage, the object query is given stronger directionality, so that it prioritizes high-frequency defect areas (such as template seams), thereby reducing invalid calculations and improving detection efficiency.

[0034] (4) The present invention introduces a hybrid loss function in the model training process, combining the classification cross entropy loss and the GIoU positioning loss. Through weight distribution, the model prioritizes learning the sensitive features of high-risk defects such as cracks, and balances multi-task training to avoid a single task dominating the optimization direction. In addition, the hyperparameters are adjustable, so the engineering adaptability is strong, and the parameters are optimized and learned during the training process.

[0035] (5) The present invention integrates deep learning technology with visual detection methods to achieve high-precision automatic recognition and quality assessment of concrete surface defects (such as cracks, honeycombs, holes, etc.).

[0036] (6) The present invention uses the end-to-end target detection framework of DETR to design a concrete pouring vibration appearance detection model, combines the self-attention mechanism with parallel decoding technology, eliminates the limitations of anchor frame design in traditional methods, and significantly improves the efficiency and accuracy of defect detection in complex scenarios.

[0037] (7) The present invention uses deep learning technology to achieve automated and high-precision quality assessment, significantly improving detection efficiency and consistency. It is suitable for various concrete pouring scenarios such as high arch dam warehouse surfaces and building floor slabs. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a structural diagram of the concrete pouring and vibration surface detection model in the present invention. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings.

[0040] The present invention proposes a method for evaluating the apparent quality of concrete pouring and vibration based on DETR, which specifically includes the following steps:

[0041] S1: Collect concrete surface images.

[0042] The present invention can use an industrial camera (e.g., 50 megapixels) in conjunction with a ring-shaped fill light to capture images of the concrete surface, ensuring image clarity and illumination uniformity. Furthermore, for complex curved surfaces, a 3D line laser scanner can be used to obtain three-dimensional surface topography data.

[0043] S2: Use the concrete pouring vibration surface detection model to perform semantic segmentation and defect detection on the images in the dataset.

[0044] The concrete pouring vibration appearance detection model described in the present invention can adopt the DETR model. The core of DETR is composed of the Transformer encoder-decoder structure and the bipartite matching loss function. Figure 1 As shown in the figure, the concrete pouring vibration appearance detection model structure includes the image input layer, image preprocessing layer, CNN backbone network layer, encoder, Transformer decoder, prediction head, post-processing layer, and output layer. The encoder extracts global features through the self-attention mechanism, and the decoder interacts the learnable object queries with the image features to generate prediction boxes and categories in parallel.

[0045] The image input layer is used to input images in the dataset.

[0046] The image preprocessing layer is used to normalize and augment the input image. Data augmentation includes random cropping (simulating image input from different perspectives to enhance the model's adaptability to target position and scale), Gaussian noise injection (simulating sensor noise or environmental interference to improve the model's robustness to low-quality images), and MixUp (linearly blending two images and their labels to enhance the model's generalization to mixed features) to improve the model's robustness to interference such as noise and occlusion.

[0047] The CNN backbone network layer serves as a feature extraction layer and adopts a ResNet network or an EfficientNet network to extract multi-scale features of the preprocessed image, including cracks, honeycombs, holes, etc., output a feature map, and provide semantic information.

[0048] A Feature Pyramid Network (FPN) is introduced at the encoder front end (designed to address the detection of objects of varying scales). This network fuses convolutional features from different levels (deep semantic information and shallow detail information) to achieve multi-scale feature fusion and improve the sensitivity of small defect detection. The FPN fuses multi-scale features, aligns semantic information from different levels through horizontal connections, and restores the spatial resolution of deep features through top-down upsampling. The feature pyramid output supports multi-scale prediction (simultaneous detection of objects of varying sizes).

[0049] The encoder consists of a position encoding layer and a Transformer encoder. The position encoding layer acts as an embedding layer, encoding the feature map positionally (position encoding: superimposing sine / cosine position encodings onto the feature map) to capture global dependencies. The Transformer encoder processes the position-encoded sequence data using a multi-head self-attention mechanism and a feedforward network (FFN), outputting global semantics and position information to the Transformer decoder.

[0050] The Transformer decoder performs object queries on the Transformer encoder output features, uses a cross-attention mechanism to interact the object queries with the encoder output features (global semantics and position information), locates the locations of concrete pouring and vibration defects, and uses a self-attention mechanism to directly interact the object queries with each other, learns and distinguishes different concrete pouring and vibration defect categories, and outputs the concrete pouring and vibration defect categories and locations in the form of feature sequences.

[0051] The prediction head is used to classify and regress the feature sequence output by the Transforme decoder, and output the category probability and location coordinates of the concrete pouring and vibration defects.

[0052] The main function of the post-processing layer is to optimize the prediction results, compare the set confidence threshold with the category probability of concrete pouring and vibration defects, and filter out the prediction head output results below the confidence threshold;

[0053] The main function of the output layer is to display the structured or visualized results, and to output the post-processed prediction results, including the category, location, and confidence level of the concrete pouring and vibration defects, where the confidence level is the category probability level.

[0054] S3: Quantitatively score the apparent quality of concrete pouring and vibration based on the defect detection results.

[0055] In the above-mentioned quality assessment method, preferably, before training the concrete pouring and vibration surface inspection model, an industrial camera (e.g., 50 megapixels) equipped with a macro lens and a ring light is used to capture concrete surface images, ensuring pixel resolution sufficient to capture cracks at the 0.1 mm level. After acquiring the surface images, they are processed and annotated according to a standard annotation format to construct a dataset containing multi-scale concrete surface images for use as the training dataset.

[0056] Annotation specifications: The collected concrete surface images must be annotated before they can be used as training datasets. Defect annotation follows the COCO format. Annotation categories include cracks (aspect ratio > 5:1), honeycombs (irregular hole groups), holes (isolated circular defects), etc.

[0057] Different processing measures are required for different complex surfaces:

[0058] Cracks on concrete surfaces are long and narrow, with widely varying widths, making them susceptible to interference from surface texture. Therefore, detection metrics include: length L; width W; length error l; width error w; processing speed; and anti-interference capability. The minimum detectable crack size is 0.1mm wide. Furthermore, in addition to length and maximum width, crack labels should also include a width marker every 10cm along the length of the crack, along with the start and end coordinates and direction angle.

[0059] Honeycombs on concrete surfaces appear as irregular clusters of holes with blurred boundaries, necessitating a distinction between densely packed holes and localized pitting. Detection metrics include minimum hole diameter D, depth estimation accuracy q, and distribution density calculation. The minimum detectable honeycomb is 5mm in diameter.

[0060] Holes on concrete surfaces vary in shape and have sharp edges, making it necessary to distinguish between surface holes and penetrating holes. Therefore, detection criteria include shape adaptability, depth classification, minimum area detection, edge accuracy, and multi-scale detection. The minimum detectable hole size is 10 mm².

[0061] In the above-mentioned quality evaluation method, preferably, during the training process of the concrete pouring vibration surface detection model, the Transformer decoder performs dynamic query initialization before performing object query, queries the training image through engineering prior knowledge, and gives the object query stronger directionality in the initialization stage, so that it gives priority to high-frequency defect areas (such as template joints), thereby reducing invalid calculations and improving detection efficiency.

[0062] In the above-mentioned quality assessment method, the loss function used in the training of the concrete pouring and vibration surface inspection model is preferably a hybrid loss function, which combines classification cross entropy loss with GIoU positioning loss and introduces defect severity weights to optimize the model's focus on key defects. Through weight distribution, the hybrid loss function prioritizes learning sensitive features of high-risk defects such as cracks, and balances multi-task training to avoid a single task dominating the optimization direction. Furthermore, adjustable hyperparameters provide strong engineering adaptability, while parameter optimization and learning are performed during the training process.

[0063] For example, the training parameters for a concrete pouring vibration surface detection model might be: batch size 16, initial learning rate 1e-4, weight decay 5e-4, and 200 training epochs. Early stopping is used to prevent overfitting. The trained model is then integrated into an edge computing device (such as the Jetson AGX Xavier) to enable real-time detection and visualization of results at the construction site.

[0064] Experimental Verification and Comparative Analysis: 2,000 concrete surface images were collected as a test set. The experiments demonstrated that this method significantly outperformed traditional models such as Faster R-CNN and YOLOv5 in defect detection accuracy (mAP@0.5 reached 92.3%) and inference speed (single-frame time ≤ 0.2s). It also demonstrated strong robustness to illumination variations and surface stains.

[0065] Based on the principle of the above-mentioned concrete pouring and vibrating surface quality evaluation method, the present invention also proposes a concrete pouring and vibrating surface quality evaluation system based on DETR, which mainly includes an image acquisition module, a defect detection module, and a quality evaluation module. The image acquisition module acquires concrete surface images, processes and annotates them according to the annotation specification format, and constructs a data set containing multi-scale concrete surface images; the defect detection module uses the concrete pouring and vibrating surface quality detection model to perform semantic segmentation and defect detection on the images in the data set; the quality evaluation module quantitatively scores the concrete pouring and vibrating surface quality based on the defect detection results (i.e., defect location, category, and confidence), and outputs a visual report.

[0066] Although the functions and working processes of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific functions and working processes. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all protected by the present invention.

Claims

1. A method for evaluating the apparent quality of concrete pouring and vibration based on DETR, characterized in that: The following steps are involved: S1: collect concrete surface images; S2: Use the concrete pouring and vibrating surface detection model to perform semantic segmentation and defect detection on the images in the dataset; S3: Quantitatively score the apparent quality of concrete pouring and vibration based on the defect detection results.

2. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 1, characterized in that: In step S1, an industrial camera is used in conjunction with a ring light to capture an image of the concrete surface.

3. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 1, characterized in that: The concrete pouring vibration appearance detection model described in step S2 adopts the DETR model, and the model structure includes an image input layer, an image preprocessing layer, a CNN backbone network layer, an encoder, a Transformer decoder, a prediction head, a post-processing layer, and an output layer, which are arranged in sequence.

4. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 3, characterized in that: The image input layer is used to input images in the data set; The image preprocessing layer is used to perform normalization and data enhancement on the input image; The CNN backbone network layer uses a ResNet network or an EfficientNet network to extract multi-scale features from the preprocessed image, including cracks, honeycombs, and holes, output feature maps, and provide semantic information; The encoder includes a position encoding layer and a Transformer encoder; The position encoding layer is used to perform position encoding on the feature map to obtain global dependencies; the Transformer encoder processes the position-encoded sequence data through a multi-head self-attention mechanism and a feedforward network, and outputs global semantics and position information to the Transformer decoder; The Transformer decoder performs object queries on the Transformer encoder output features, uses a cross-attention mechanism to interact the object queries with the encoder output features, locates the location of concrete pouring and vibration defects, and uses a self-attention mechanism to interact the object queries directly with each other, learns and distinguishes different concrete pouring and vibration defect categories, and outputs the concrete pouring and vibration defect categories and locations in the form of feature sequences; The prediction head is used to classify and regress the feature sequence output by the Transforme decoder, and output the category probability and location coordinates of the concrete pouring and vibration defects; The post-processing layer is used to compare the set confidence threshold with the category probability of concrete pouring and vibration defects, and filter out the prediction head output results below the confidence threshold; The output layer is used to output the post-processed prediction results, including the category, location, and confidence level of the concrete pouring and vibration defects.

5. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 3, characterized in that: The training dataset acquisition process of the concrete pouring and vibrating surface detection model is as follows: an industrial camera is used in conjunction with a ring fill light to capture concrete surface images, which are processed and annotated according to the annotation standard format to construct a dataset containing multi-scale concrete surface images; Annotation specification: The collected concrete surface images follow the COCO format for defect annotation, and the annotation categories include cracks, honeycombs, and holes; For cracks on the concrete surface, the detection indicators include: length, width, length error, width error, processing speed and anti-interference ability; for honeycombs on the concrete surface, the detection indicators include: minimum hole diameter and depth estimation accuracy; for holes on the concrete surface, the detection indicators include: shape adaptability, depth classification, minimum area detection, edge accuracy, and multi-scale detection.

6. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 3, characterized in that: The image preprocessing layer performs data enhancement on the input image, including random cropping, Gaussian noise injection and MixUp enhancement.

7. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 3, characterized in that: A feature pyramid network is introduced at the front end of the encoder to fuse convolutional features at different levels and realize multi-scale feature fusion.

8. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 3, characterized in that: During the training process of the concrete pouring and vibration surface detection model, the Transformer decoder performs dynamic query initialization before performing object query, queries the training image based on engineering prior knowledge, and finds high-frequency defect areas.

9. The method for evaluating the apparent quality of concrete pouring and vibration based on DETR according to claim 3, characterized in that: The loss function used in the training process of the concrete pouring vibration appearance detection model is a hybrid loss function that combines classification cross entropy loss and GIoU positioning loss.

10. A concrete pouring and vibrating apparent quality evaluation system based on DETR, characterized in that: It includes an image acquisition module for acquiring images of the concrete surface; The defect detection module uses a concrete pouring and vibration surface detection model to perform semantic segmentation and defect detection on images in the dataset; The quality evaluation module quantitatively scores the apparent quality of concrete pouring and vibration based on the defect detection results.

Images