Road disease identification method, system, device and medium based on semi-supervised learning

Abstract

The present application relates to the technical field of computer vision and road detection, and particularly relates to a pavement disease identification method, system, device and medium based on semi-supervised learning. The method comprises: processing a pavement image through an image recognition model to obtain an identification result containing disease category and spatial position information; based on the identification result, quantifying a disease area according to a grid of a preset size to obtain a quantification result; based on the identification result and the quantification result, generating a structured data file; in response to a review operation on disease data in the structured data file, correcting the disease data based on a user interaction instruction to obtain corrected disease data; and based on a data set containing the corrected disease data, iteratively training the image recognition model. Compared with the prior art, the present application effectively closes the loop between the manual review and correction link and the model training link, so that the model can be iteratively optimized using continuously generated correction data to realize self-evolution of performance.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and road inspection technology, and in particular to a method, system, device and medium for road surface defect identification based on semi-supervised learning. Background Technology

[0002] Automatic identification of road surface defects is a core technology in the field of intelligent highway inspection and is crucial for road maintenance decisions. Currently, deep learning-based image recognition models are widely used for defect detection and segmentation of road surface images. However, existing technologies typically treat model identification and the verification and correction of labeled data as two separate processes. After the model completes identification and quantization, its output is mostly used to generate inspection reports. Even if manual corrections are made to the identification results during the verification process, these high-quality correction data are not effectively used to retrain the model. This results in the model being unable to learn from errors in real-world applications, leading to stagnant recognition performance and difficulty in adapting to changes in different road sections, seasons, or imaging conditions.

[0003] Existing technology CN120953801A discloses a bridge defect identification and maintenance method based on a multimodal large model, which trains a damage detection and segmentation model by constructing a labeled dataset and finally uses a large language model to generate a maintenance plan; Existing technology CN120877121A discloses a highway defect detection method that can accurately locate defects, which achieves accurate defect identification and geographic information binding by optimizing the instance segmentation model and combining multi-target tracking and GPS positioning.

[0004] However, the aforementioned and similar existing technical solutions generally suffer from a key flaw: they separate the automatic identification process of the model from the subsequent manual review and correction of the identification results. After the model completes identification and quantification, the output is mostly used to generate inspection reports, and the process ends there. Even if maintenance experts perform high-quality manual corrections to the model's identification errors or omissions during report review, this corrective data is only used as the basis for finalizing the report and is not systematically collected and fed back into the model training process. This results in the model always being trained on the initial static dataset, unable to learn from the continuously generated correction signals in real-world application scenarios. Its recognition performance stagnates long after deployment, making it difficult to adapt to changes in data distribution caused by different road materials, diverse environmental conditions (such as seasonal and lighting changes), and differences in imaging equipment. Therefore, the errors generated by the model in practical applications will be repeated, preventing self-optimization and evolution.

[0005] Therefore, there is an urgent need for a technical solution that can effectively close the loop between the manual review and correction process and the model training process, so that the model can use the continuously generated correction data to perform iterative optimization and achieve self-evolution of performance. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a road surface defect identification scheme based on semi-supervised learning, so as to overcome the technical defects of the prior art that the model cannot be continuously optimized by manually reviewing and correcting data, resulting in stagnant identification performance and weak adaptive ability.

[0007] The objective of this invention can be achieved through the following technical solutions: To address the aforementioned technical problems, this invention provides a road surface defect identification method based on semi-supervised learning. The method includes acquiring a road surface image; processing the road surface image using an image recognition model to obtain a recognition result containing defect category and spatial location information; quantizing the defect area according to a grid of preset size based on the recognition result to obtain a quantization result; generating a structured data file based on the recognition result and the quantization result; correcting the defect data based on user interaction commands in response to a review operation of the defect data in the structured data file to obtain corrected defect data; and iteratively training the image recognition model based on a dataset containing the corrected defect data.

[0008] Further, the step of processing the road surface image using an image recognition model to obtain a recognition result containing information on the type and spatial location of the defects includes, for the first type of defect, calling a first segmentation model to process the road surface image to obtain a first segmentation result. For the second type of defect, calling a target detection model to process the road surface image to obtain the location of the defect area, and based on the location of the defect area, calling the corresponding second segmentation model to process it to obtain a second segmentation result. Based on the first segmentation result and the second segmentation result, the recognition result is generated.

[0009] In one implementation, based on the identification result, the damaged area is quantized according to a grid of preset size to obtain a quantization result. This includes determining the mapping relationship between pixel size and physical size based on the resolution of the road image and camera calibration parameters. Based on the mapping relationship and the preset size, a grid is superimposed on the damaged area corresponding to the identification result. Geometric parameters of the damaged area are calculated based on the grid to obtain the quantization result.

[0010] Further, the step of calculating geometric parameters of the diseased area based on the grid to obtain the quantization result includes, in response to the disease category belonging to a first quantization rule category, calculating a first area value based on the grid and using the first area value as the area parameter in the quantization result. In response to the disease category belonging to a second quantization rule category, calculating a second area value based on the grid and using the second area value as the area parameter in the quantization result. The first area value is twice the second area value.

[0011] In one embodiment, the structured data file is a JSON file, which includes at least one of the following: image index field, disease category field, mask coordinate field, area parameter field, length parameter field, grid ratio information field, and identification confidence field.

[0012] Furthermore, in response to the review operation of the defect data in the structured data file, the defect data is corrected based on user interaction instructions to obtain corrected defect data. This includes loading and displaying the pavement image, the recognition result, and the quantization result corresponding to the structured data file in the visualization interface. In response to the received user interaction instructions, an interactive segmentation model is invoked to generate a supplementary segmentation mask. Based on the supplementary segmentation mask, the defect data in the structured data file is updated to obtain the corrected defect data.

[0013] As one implementation, the user interaction instructions include at least one of text instructions based on natural language description, coordinate instructions based on box selection, and point coordinate instructions based on point selection.

[0014] Furthermore, the step of invoking the interactive segmentation model to generate a supplementary segmentation mask in response to a received user interaction instruction includes, if the user interaction instruction is a text instruction based on natural language description, encoding the text instruction into a text feature vector. The text feature vector is then input into the interactive segmentation model to generate the supplementary segmentation mask.

[0015] Furthermore, the iterative training of the image recognition model based on the dataset containing the corrected disease data includes introducing adaptation parameters into the image recognition model and freezing the basic parameters of the image recognition model other than the adaptation parameters. The adaptation parameters are then trained based on the dataset.

[0016] In one embodiment, introducing adaptation parameters into the image recognition model includes performing low-rank decomposition on the weight matrix of at least one fully connected layer in the image recognition model to obtain a low-rank adaptation matrix. The low-rank adaptation matrix is ​​then used as the adaptation parameter.

[0017] Furthermore, the dataset includes a first data subset and a second data subset, wherein the first data subset contains the corrected disease data, and the second data subset contains labeled data automatically generated by the image recognition model during training. Training the adaptation parameters based on the dataset includes jointly training the adaptation parameters based on the first data subset and the second data subset.

[0018] This invention also provides a road surface defect identification system, comprising: an image acquisition module for acquiring road surface images; a defect identification module for processing the road surface images using an image recognition model to obtain identification results containing defect category and spatial location information; a defect quantization module for quantizing defect areas based on the identification results according to a grid of preset size to obtain quantized results; a data generation module for generating a structured data file based on the identification results and the quantization results; a verification and correction module for correcting the defect data based on user interaction commands in response to a verification operation of the defect data in the structured data file to obtain corrected defect data; and a model training module for iteratively training the image recognition model based on a dataset containing the corrected defect data.

[0019] Compared with the closest prior art, the present invention has the following advantages: This invention provides a standardized data foundation for efficient human-computer interaction review by generating structured data files containing recognition and quantification results. During the review process, users can correct the defect data based on various interactive commands, and the corrected defect data is directly used to construct the training dataset. By iteratively training the image recognition model based on this training set containing high-quality manually corrected data, the model can continuously learn from errors and feedback in practical applications, effectively solving the problem of model performance stagnation in existing technologies and achieving closed-loop self-evolution of the model. Specifically, the structured data file (such as a JSON file) allows defect information to be stored in a way that can be conveniently processed by both machines and humans, reducing the operational threshold for review and correction. By introducing an interactive segmentation model to generate supplementary segmentation masks in response to user commands, high-precision corrected data can be generated quickly. During the iterative training phase, by introducing and training only adaptation parameters (such as a low-rank adaptation matrix), the model can quickly absorb new knowledge while greatly reducing training computational overhead and avoiding catastrophic forgetting, thereby continuously improving the recognition accuracy and generalization ability of road defects in long-term applications. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a semi-supervised learning-based method for identifying road surface defects, as provided in an embodiment of this application.

[0021] Figure 2 This is a schematic diagram illustrating the overlay of a pre-sized grid on the diseased area, as provided in an embodiment of this application.

[0022] Figure 3 This is a schematic diagram of an interactive segmentation interface based on box selection provided in an embodiment of this application.

[0023] Figure 4This is a schematic diagram of a point-based interactive segmentation interface provided in an embodiment of this application.

[0024] Figure 5 A schematic diagram of an interactive segmentation interface based on natural language provided in an embodiment of this application.

[0025] Figure 6 The mAP training curves for three rounds of closed-loop iterations are used to verify the improvement of model accuracy through iteration.

[0026] Figure 7 The training curve for mAP@50 is obtained through three rounds of closed-loop iteration.

[0027] Figure 8 This is a graph showing the relationship between the optimal mAP and data size across three iterations.

[0028] Figure 9 To compare the three-round iterative model mAP. Detailed Implementation

[0029] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, circuit structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0030] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0031] Example 1 The pavement defect identification method provided in this application can be applied to defect identification and assessment scenarios for highway pavements, urban road pavements, or airport runways. This method integrates image recognition, defect geometric quantification, human-computer interaction verification, and iterative model training into a closed-loop system. It can utilize continuously generated correction data to drive the self-evolution of model performance, effectively solving the technical problems of stagnant model performance and weak adaptive capabilities in traditional methods.

[0032] Figure 1 A flowchart illustrating a semi-supervised learning-based method for pavement defect identification is shown. Figure 1As shown, embodiments of this application provide a method for identifying road surface defects based on semi-supervised learning. The method includes the following steps: S1: acquiring a road surface image; S2: processing the road surface image using an image recognition model to obtain a recognition result containing defect category and spatial location information; S3: quantizing the defect area based on the recognition result using a grid of preset size to obtain a quantization result; S4: generating a structured data file based on the recognition result and the quantization result; S5: in response to a verification operation on the defect data in the structured data file, correcting the defect data based on user interaction instructions to obtain corrected defect data; and S6: iteratively training the image recognition model based on a dataset containing the corrected defect data. This method forms a closed loop between identification, quantization, verification / correction, and model retraining, enabling the model to continuously learn from manually corrected high-quality data, thereby continuously improving recognition accuracy and generalization ability.

[0033] In this embodiment, step S2 can be specifically executed as step S21. When executing step S21, it includes: for the first type of defect, calling the first segmentation model to process the road surface image to obtain a first segmentation result; and for the second type of defect, calling the target detection model to process the road surface image to obtain the defect area location, and calling the corresponding second segmentation model based on the defect area location to obtain a second segmentation result; finally, generating a recognition result based on the first and second segmentation results. The image recognition model here is actually a multi-model collaborative system. The first type of defect may refer to defects with relatively simple morphological features that are easy to directly segment (e.g., linear cracks), so a pre-trained first segmentation model (such as a dedicated model based on U-Net) can be directly used for pixel-level recognition. The second type of disease may refer to diseases with complex shapes or low contrast with the background (such as pits or loose areas). First, a target detection model (such as a YOLO or DINO-based model) is used to locate its approximate area, which narrows the processing range, reduces computational complexity, and improves accuracy. Then, a second segmentation model trained specifically for this type of disease is called for fine segmentation, thus balancing recognition efficiency and accuracy. The processor loads the weight parameters of these pre-trained models into memory and performs forward propagation calculations, ultimately outputting recognition results containing disease category labels, segmentation mask coordinates, or bounding boxes.

[0034] In this embodiment, step S3 can be specifically executed as step S31.

[0035] When step S31 is executed, it includes: determining the mapping relationship between pixel size and physical size based on the resolution of the road image and camera calibration parameters; overlaying a grid onto the defect area corresponding to the identification result based on the mapping relationship and a preset size; and calculating the geometric parameters of the defect area based on the grid to obtain the quantization result. The processor calculates the actual physical size (e.g., centimeters / pixel) corresponding to each pixel based on known camera intrinsic parameters (focal length, principal point) and extrinsic parameters, combined with image resolution, through geometric projection relationships. Subsequently, according to this mapping relationship and a preset physical size (e.g., 10cm), virtual grid lines are generated in the image coordinate system and overlaid and rendered onto the defect mask area, such as... Figure 2 As shown. This process is implemented in the processor through matrix operations and graphics drawing instructions. Further, the geometric parameter calculation can be executed as step S311. When executing step S311, it includes: in response to the disease category belonging to a first quantization rule category, calculating a first area value based on the grid, and using the first area value as the area parameter in the quantization result; in response to the disease category belonging to a second quantization rule category, calculating a second area value based on the grid, and using the second area value as the area parameter in the quantization result; wherein, the first area value is twice the second area value. For example, the first quantization rule category may correspond to linear diseases (such as longitudinal cracks and transverse cracks) that are required to be calculated based on the width of influence in highway specifications, and its area is calculated as twice the total area of ​​the covering grid; the second quantization rule category may correspond to diseases calculated based on the actual area (such as potholes, network cracks, and loose areas), and its area is calculated based on the total area of ​​the covering grid. The processor calculates the final area parameter by traversing the grid cells, counting the number of cells covered by the disease mask, and then multiplying it by the actual area of ​​each cell (or twice the actual area of ​​the cell), and writes the result into the quantization result data structure in memory.

[0036] In this embodiment, the structured data file generated in step S4 is preferably a JSON file. This JSON file includes at least one of the following: image index field, disease category field, mask coordinate field, area parameter field, length parameter field, grid ratio information field, and identification confidence field. The processor serializes the identification and quantization results generated in steps S2 and S3 into a JSON string according to a predefined key-value pair format and saves it to storage (such as a hard drive or database). This structured format facilitates subsequent machine reading, parsing, and manual visual verification.

[0037] In this embodiment, step S5 can be specifically executed as step S51.

[0038] When step S51 is executed, it includes: loading and displaying the road image, the recognition result, and the quantization result corresponding to the structured data file in the visualization interface; in response to the received user interaction command, calling the interactive segmentation model to generate a supplementary segmentation mask; and updating the defect data in the structured data file based on the supplementary segmentation mask to obtain the corrected defect data.

[0039] like Figures 3 to 5 As shown, the visualization interface is rendered by a graphical user interface (GUI) program that reads a JSON file and the original image from memory and overlays lesion masks, category labels, and quantization grids. User interaction commands include at least one of the following: text commands based on natural language descriptions, coordinate commands based on bounding box operations, and point coordinate commands based on point selection operations. The processor captures these commands by listening for events from input devices such as the mouse and keyboard.

[0040] Specifically, in response to the user interaction instruction being a text instruction based on natural language description, step S51 further includes: encoding the text instruction into a text feature vector; inputting the text feature vector into the interactive segmentation model to generate the supplementary segmentation mask. This can be achieved by integrating a vision-language model (such as Segment AnythingModel, SAM). The processor uses a pre-trained text encoder (such as CLIP's text encoder) to convert the user-input natural language (such as "the crack in the lower left corner") into a high-dimensional feature vector, and then inputs this vector along with the cached visual features of the original image into the decoder of the interactive segmentation model to quickly generate the supplementary segmentation mask. The generated mask is immediately overlaid and displayed. After user confirmation, the processor calculates its quantization parameters, updates the corresponding fields in the JSON file, and generates a new copy of the file with a version identifier, which is then stored in memory.

[0041] In this embodiment, step S6 can be specifically executed as step S61. When executing step S61, it includes: introducing adaptation parameters into the image recognition model and freezing the basic parameters of the image recognition model other than the adaptation parameters; training the adaptation parameters based on the dataset. This typically employs Parameter Efficient Fine-Tuning (PEFT) techniques, such as Low-Rank Adaptation (LoRA). Specifically, introducing adaptation parameters can be executed as step S611: performing low-rank decomposition on the weight matrix of at least one fully connected layer in the image recognition model to obtain a low-rank adaptation matrix; using the low-rank adaptation matrix as the adaptation parameters. During training, the processor freezes all basic parameters of the original model and only performs gradient descent optimization on the newly added low-rank adaptation matrix, which greatly reduces the computational resources and time required for training and effectively avoids catastrophic forgetting. The dataset includes a first data subset and a second data subset, the first data subset containing the corrected disease data, and the second data subset containing labeled data (i.e., pseudo-labels) automatically generated by the image recognition model during training. The adaptation parameters are trained based on the dataset, specifically by jointly training the adaptation parameters based on the first data subset and the second data subset. The processor loads the data from these two subsets from memory, performs mixed sampling in each training iteration, calculates the loss (such as segmentation loss and detection loss), and backpropagates to update the weights of the low-rank adaptation matrix.

[0042] Verification Example 1 Actual road tests show that this method, through the aforementioned closed-loop iterative mechanism, continuously improves model performance. Using only initial manually labeled data for fully supervised training as the baseline (iteration 0, average accuracy mAP = 0.287), after the first round of retraining using model-generated pseudo-labels and subsequent verification and correction (iteration 1), mAP improved to 0.406 (a relative improvement of 41.5%). After the second round of closed-loop iteration (iteration 2), mAP further improved to 0.411 (a relative improvement of 43.2% compared to the baseline). Figure 4 As shown, under the same training rounds, the model performance after each iteration is systematically higher than that of the previous round, which fully verifies that the closed-loop semi-supervised mechanism of this invention has a truly quantifiable self-evolutionary capability.

[0043] Specifically, the experiment used a professional dataset covering six common pavement defects: longitudinal cracks, mesh cracks, sealing strips, transverse cracks, potholes, and repairs. The entire validation process included three complete training iterations to simulate the closed-loop mechanism described in this invention. Iteration 0 served as the performance baseline. Referring to Table 1, the DINO detection model was trained in a fully supervised manner using only the initial 16,470 manually labeled images, achieving an optimal mean average precision (mAP) of 0.287. This simulates the state in traditional technical solutions where the model is trained on a fixed dataset and then deployed, thus solidifying its performance.

[0044] Table 1: Training Data Information from Three Rounds Detection model: DINO (Transformer-based end-to-end target detector), backbone network: Swin-L Dataset: A dedicated dataset for road defect detection, covering 6 types of defects—longitudinal cracks, mesh cracks, sealing strip defects, transverse cracks, potholes, and repairs. Evaluation metrics: COCO standard protocol, mAP (IoU threshold 0.50:0.95), mAP@50 (IoU=0.50), mAP@75 (IoU=0.75) Hardware environment: NVIDIA A800-SXM4-80GB GPU × 8 Figure 6 The chart shows a comparison of the mAP training curves for three rounds of closed-loop iteration. The three curves are arranged in layers from bottom to top, intuitively showing the trend of the overall detection accuracy continuously increasing with each training round after each round of closed-loop iteration, verifying the performance improvement effect of closed-loop iteration.

[0045] Figure 7 The chart shows a comparison of the mAP@50 training curves for three rounds of closed-loop iterations. The three curves are arranged in layers, clearly demonstrating the model's detection performance at the IoU=0.50 threshold. The performance gradually improves with the increase of the closed-loop iteration rounds, reflecting the effectiveness of the self-evolution mechanism.

[0046] Figure 8 The graph shows the relationship between the optimal mAP and the training data size over three iterations. The optimal mAP and mAP@50 for each iteration are displayed in a bar chart, and the growth of the training set size is marked on the right axis line, which intuitively reflects the positive correlation between the increase in data volume and the improvement in model performance.

[0047] Figure 9 The graph shows a hierarchical comparison of mAP of the three iterative models under the same training epochs. The comparison is made within the common training interval from Epoch 5 to 10. The three curves form a clear performance hierarchy, proving that the improvement in model performance comes from closed-loop data optimization rather than simply the accumulation of training epochs.

[0048] Starting with the first closed-loop iteration, the model trained in iteration 0 automatically identifies a large number of unlabeled images, generating preliminary pseudo-label results. These automatically identified results then enter the crucial human-machine collaborative review stage, where professionals quickly correct and confirm identification errors or omissions, thus generating high-quality corrected labeled data. This corrected data is merged with the initial labeled data to form an enhanced training set of 29,624 images, which is then used to retrain the model. Experimental results show that the optimal mAP of the iteration 1 model is significantly improved to 0.406, representing a performance improvement of up to 41.5% compared to the baseline model. This initially demonstrates that feeding back the corrected data generated in the manual review stage to the training process can effectively overcome model performance bottlenecks.

[0049] Referring to Tables 1 to 3, a second round of closed-loop iteration was conducted to further verify the sustainability of the performance improvement. The pseudo-labels were generated again using the more powerful Iteration 1 model, and after the same human-machine collaborative review and correction, the training set size was further expanded to 67,656 images. After retraining, the optimal mAP of Iteration 2 model reached 0.411, an improvement of 43.2% compared to the baseline. The optimal mAP of the three iterations showed a monotonically increasing trend of 0.287, 0.406, and 0.411, directly proving that the closed-loop mechanism of this invention can drive continuous improvement in model performance with increasing iteration rounds.

[0050] Referring to Tables 1 to 3, to exclude the benefits solely from extended training time, the training period was fixed, and the performance of each generation of models was compared under the same number of training rounds. Within the common interval of rounds 5 to 10, the mAP of iteration 1 model was systematically higher than that of iteration 0 model at every point in the same round, while iteration 2 model was systematically higher than iteration 1 model. For example, in round 10, the mAP of iteration 0, iteration 1, and iteration 2 models were 0.275, 0.334, and 0.347, respectively. This clear performance stratification under the same training input indicates that the corrected data introduced in each closed-loop round substantially improved the model's knowledge base and learning starting point, enabling the new generation model to possess stronger recognition potential than the previous generation model from the very beginning of training. This proves that the performance improvement stems from a fundamental improvement in data quality, rather than simply the accumulation of training time.

[0051] Table 2: Comparison of Optimal Detection Performance by Iteration Round Table 2a: Iteration 0 Training Process (Initial Manually Annotated Data, Ep5-14) Table 2b: Iteration 1 Training Process (Pseudo-label Augmentation Round 1, Ep1-34) Table 2c: Iteration 2 Training Process (Pseudo-label augmentation round 2, Ep1-36) Under the same training input (Epoch 5~10), the mAP of the three-round iterative model is directly compared, excluding the difference in training rounds, and the performance gain brought by pseudo-label augmentation in each round is measured separately.

[0052] Table 3: Horizontal comparison of mAP across three iterations in the common training interval (Epoch 5~10) This application also provides a pavement defect identification system, which includes an image acquisition module, a defect identification module, a defect quantification module, a data generation module, a review and correction module, and a model training module. These modules can be stored in memory as software and executed by a processor to implement the various steps in the aforementioned method embodiments. For example, the image acquisition module performs step S1, the defect identification module performs step S2, the defect quantification module performs step S3, the data generation module performs step S4, the review and correction module performs step S5, and the model training module performs step S6.

[0053] This application also provides an electronic device, which includes a processor and a memory.

[0054] Memory is used to store computer programs. Memory can be random access memory (RAM), read-only memory (ROM), etc.

[0055] The processor executes the computer program stored in the memory to implement the steps in the semi-supervised learning-based pavement defect identification method described above. The processor can be an integrated circuit chip with signal processing capabilities; for example, it can be a general-purpose processor (such as a CPU), graphics processing unit (GPU), digital signal processor (DSP), application-specific integrated circuit (ASIC), etc. The processor and memory can be connected via a bus, and the electronic device may also include conventional components such as communication interfaces and input / output devices.

[0056] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the road surface defect identification method based on semi-supervised learning in the above embodiments.

[0057] The computer-readable storage medium can be any medium capable of storing program code, such as cloud storage, USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0058] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A method for identifying pavement defects based on semi-supervised learning, characterized in that, Includes the following steps: Acquire road surface images; The road surface image is processed by an image recognition model to obtain recognition results containing information on the type and spatial location of the damage. Based on the identification results, the diseased area is quantized according to a grid of preset size to obtain the quantization result; Based on the identification results and the quantization results, a structured data file is generated; In response to the review operation of the disease data in the structured data file, the disease data is corrected based on user interaction instructions to obtain the corrected disease data; Based on the dataset containing the corrected disease data, the image recognition model is iteratively trained. The recognition model is used to identify road surface defects after reaching a preset number of training iterations.

2. The method for identifying pavement defects based on semi-supervised learning according to claim 1, characterized in that, The road surface image is processed using an image recognition model to obtain recognition results containing information on the type and spatial location of the damage, including: For the first type of damage, the first segmentation model is called to process the road surface image to obtain the first segmentation result; For the second type of damage, the target detection model is called to process the road surface image to obtain the location of the damage area, and the corresponding second segmentation model is called to process the damage area based on the location of the damage area to obtain the second segmentation result; The recognition result is generated based on the first segmentation result and the second segmentation result.

3. The method for identifying road surface defects based on semi-supervised learning according to claim 1, characterized in that, Based on the identification results, the diseased area is quantized according to a grid of preset size to obtain the quantization results, including: Based on the resolution of the road surface image and the camera calibration parameters, the mapping relationship between pixel size and physical size is determined; Based on the mapping relationship and the preset size, a grid is superimposed on the diseased area corresponding to the identification result; The geometric parameters of the diseased area are calculated based on the grid to obtain the quantification result.

4. The method for identifying pavement defects based on semi-supervised learning according to claim 3, characterized in that, Based on the grid, geometric parameters of the diseased area are calculated to obtain the quantification results, including: In response to the disease category belonging to the first quantification rule category, a first area value is calculated based on the grid, and the first area value is used as the area parameter in the quantification result; In response to the disease category belonging to the second quantification rule category, a second area value is calculated based on the grid, and the second area value is used as the area parameter in the quantification result; Wherein, the first area value is twice the second area value.

5. The method for identifying pavement defects based on semi-supervised learning according to claim 1, characterized in that, In response to the review operation of the disease data in the structured data file, the disease data is corrected based on user interaction instructions to obtain corrected disease data, including: The road surface image, the recognition result, and the quantization result corresponding to the structured data file are loaded and displayed in the visualization interface. In response to the received user interaction command, the interactive segmentation model is invoked to generate a supplementary segmentation mask; Based on the supplementary segmentation mask, the disease data in the structured data file is updated to obtain the corrected disease data; The user interaction instructions include at least one of the following: text instructions based on natural language description, coordinate instructions based on box selection, and point coordinate instructions based on point selection. The structured data file is a JSON file, which includes at least one of the following: image index field, disease category field, mask coordinate field, area parameter field, length parameter field, grid ratio information field, and identification confidence field.

6. The method for identifying pavement defects based on semi-supervised learning according to claim 5, characterized in that, In response to received user interaction commands, the interactive segmentation model is invoked to generate a supplementary segmentation mask, including: In response to the user interaction instruction being a text instruction based on natural language description, the text instruction is encoded into a text feature vector; The text feature vector is input into the interactive segmentation model to generate the supplementary segmentation mask.

7. The method for identifying pavement defects based on semi-supervised learning according to claim 5, characterized in that, Based on a dataset containing the corrected disease data, the image recognition model is iteratively trained, including: An adaptation parameter is introduced into the image recognition model, and the basic parameters of the image recognition model other than the adaptation parameter are frozen. The adaptation parameters are trained based on the dataset. The introduction of adaptation parameters into the image recognition model includes: The weight matrix of at least one fully connected layer in the image recognition model is decomposed into a low-rank matrix to obtain a low-rank adaptation matrix. The low-rank adaptation matrix is ​​used as the adaptation parameter; The dataset includes a first data subset and a second data subset, wherein the first data subset contains the corrected disease data and the second data subset contains labeled data automatically generated by the image recognition model during the training process; The step of training the adaptation parameters based on the dataset includes: jointly training the adaptation parameters based on the first data subset and the second data subset.

8. A road surface defect identification system, characterized in that, include: The image acquisition module is used to acquire images of the road surface. The road surface image recognition module is used to process the road surface image through an image recognition model to obtain recognition results containing information on the type and spatial location of the road surface defects. The disease quantification module is used to quantify the disease area according to a grid of preset size based on the identification results, and obtain the quantification results; The data generation module is used to generate a structured data file based on the identification result and the quantization result; The review and correction module is used to respond to the review operation of the disease data in the structured data file, and correct the disease data based on the user interaction command to obtain the corrected disease data. The model training module is used to iteratively train the image recognition model based on a dataset containing the corrected disease data.

9. An electronic device, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images