Road crack detection method and device

Abstract

The invention discloses a road crack detection method and device applied to the field of image processing, and the method comprises the steps: obtaining a to-be-detected road image; inputting the to-be-detected road image into the target detection model to obtain a crack detection result; the crack detection result at least comprises a crack position and a crack type in the to-be-detected road image; the target detection model is obtained by performing incremental training on the initial target detection model by using a cosine annealing algorithm; the initial target detection model is obtained based on pre-training of the sample road image. According to the method and the device, through pre-training and incremental training, the model can adapt to road disease crack detection tasks under different environments and conditions, so that the detection efficiency is improved, and the road crack can be accurately identified.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular, to a road crack detection method and device. Background Art

[0002] The timely and accurate detection of road disease cracks is crucial for ensuring road safety, extending service life, and reducing maintenance costs. With the acceleration of the urbanization process and the increase in traffic volume, road crack and disease problems are becoming increasingly serious, which not only affect driving safety but also increase the burden of road maintenance.

[0003] Currently, the detection of road disease cracks mainly relies on manual inspections. Inspectors conduct inspections by walking along the road, visually observing cracks, taking photos, and recording data. However, the detection range that inspectors can complete per day is usually only 8 - 10 kilometers, with low efficiency. Moreover, due to the subjectivity of manual detection, it is easy to miss detections or make misjudgments, and the miss detection rate is as high as 25% - 35%.

[0004] Therefore, how to ensure the accurate detection of road cracks while improving the detection efficiency has become a difficult problem to be solved urgently. Summary of the Invention

[0005] This application provides a road crack detection method and device, aiming to ensure the accurate detection of road cracks while improving the detection efficiency.

[0006] To achieve the above objective, this application provides the following technical solutions:

[0007] A road crack detection method, comprising:

[0008] Obtain a road image to be detected;

[0009] Input the road image to be detected into a target detection model to obtain a crack detection result; the crack detection result at least includes the crack position and crack type in the road image to be detected; the target detection model is obtained by incrementally training an initial target detection model using the cosine annealing algorithm; the initial target detection model is pre - trained based on sample road images.

[0010] Optionally, the step of inputting the road image to be detected into the target detection model to obtain a crack detection result includes:

[0011] Obtain a query vector in the target detection model;

[0012] Input the road image to be detected into the encoder in the target detection model to obtain road image features;

[0013] Input the road image features and the query vector into the decoder in the target detection model to obtain the crack detection result.

[0014] Optionally, the process of pre-training the initial target detection model based on sample road images includes:

[0015] Obtain sample road images;

[0016] Input the sample road images into the initial target detection model to obtain sample crack detection results;

[0017] Calculate the loss function between the sample crack detection results and the actual crack detection results corresponding to the sample road images; the loss function includes at least the cross-entropy loss function; the actual crack detection results include crack composite labels; the crack composite labels are composed of local annotation boxes and global annotation boxes; the global annotation box indicates a box containing the entire length of the crack; the local annotation box contains a box with a partial length of the crack;

[0018] When the loss function does not converge, use the Hungarian matching algorithm to adjust the model parameters of the initial target detection model, and return to execute the step of inputting the sample road images into the initial target detection model to obtain sample crack detection results;

[0019] When the loss function converges, determine that the training of the initial target detection model is completed.

[0020] Optionally, the obtaining of the sample road images includes:

[0021] Obtain initial road images;

[0022] Perform normalization processing on the initial road images to obtain the processed initial road images;

[0023] Determine the processed initial road images as sample road images.

[0024] Optionally, the process of incrementally training the initial target detection model using the cosine annealing algorithm to obtain the target detection model includes:

[0025] Obtain incremental sample road images;

[0026] Freeze a preset number of layers in the backbone network of the target detection model;

[0027] After the freezing of the preset number of layers is completed, input the incremental sample road images into the target detection model to obtain incremental crack detection results;

[0028] Calculate the loss function between the incremental crack detection result and the actual crack detection result corresponding to the incremental sample road image;

[0029] When the loss function does not converge, adjust the learning rate using the cosine annealing algorithm, and return to execute the step of inputting the incremental sample road image into the target detection model to obtain the incremental crack detection result;

[0030] When the loss function converges, determine the initial target detection model as the target detection model.

[0031] A road crack detection device, comprising:

[0032] An acquisition unit for acquiring a road image to be detected;

[0033] A detection unit for inputting the road image to be detected into a target detection model to obtain a crack detection result; the crack detection result at least includes the crack position and crack type in the road image to be detected; the target detection model is obtained by incrementally training an initial target detection model using the cosine annealing algorithm; the initial target detection model is pre-trained based on sample road images.

[0034] Optionally, the detection unit is specifically configured to:

[0035] Obtain a query vector in the target detection model;

[0036] Input the road image to be detected into the encoder in the target detection model to obtain road image features;

[0037] Input the road image features and the query vector into the decoder in the target detection model to obtain a crack detection result.

[0038] Optionally, the detection unit includes:

[0039] An image acquisition unit for acquiring a sample road image;

[0040] An input unit for inputting the sample road image into the initial target detection model to obtain a sample crack detection result;

[0041] A calculation unit for calculating the loss function between the sample crack detection result and the actual crack detection result corresponding to the sample road image; the loss function at least includes a cross-entropy loss function; the actual crack detection result contains a crack composite label; the crack composite label is composed of a local annotation box and a global annotation box; the global annotation box indicates a box containing the entire length of the crack; the local annotation box contains a box with a partial length of the crack;

[0042] A return unit, configured to, when the loss function does not converge, adjust the model parameters of the initial target detection model by using the Hungarian matching algorithm, and return to execute the step of inputting the sample road image into the initial target detection model to obtain a sample crack detection result;

[0043] A determination unit, configured to, when the loss function converges, determine that the training of the initial target detection model is completed.

[0044] Optionally, the image acquisition unit is specifically configured to:

[0045] Acquire an initial road image;

[0046] Perform normalization processing on the initial road image to obtain a processed initial road image;

[0047] Determine the processed initial road image as a sample road image.

[0048] Optionally, the detection unit is specifically configured to:

[0049] Acquire an incremental sample road image;

[0050] Freeze a preset number of layers of the backbone network in the target detection model;

[0051] After the preset number of layers is frozen, input the incremental sample road image into the target detection model to obtain an incremental crack detection result;

[0052] Calculate a loss function between the incremental crack detection result and the actual crack detection result corresponding to the incremental sample road image;

[0053] When the loss function does not converge, adjust the learning rate by using the cosine annealing algorithm, and return to execute the step of inputting the incremental sample road image into the target detection model to obtain an incremental crack detection result;

[0054] When the loss function converges, determine the initial target detection model as the target detection model.

[0055] The technical solution provided by the present application acquires a road image to be detected; inputs the road image to be detected into a target detection model to obtain a crack detection result; the crack detection result at least includes the crack position and crack type in the road image to be detected; the target detection model is obtained by incrementally training an initial target detection model by using the cosine annealing algorithm; the initial target detection model is pre-trained based on a sample road image. In the present application, through the pre-training and incremental training methods, the model can adapt to road disease crack detection tasks in different environments and conditions, which not only improves the detection efficiency but also ensures accurate identification of road cracks. Brief Description of the Drawings

[0056] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0057] Figure 1 It is a flowchart of a road crack detection method provided by an embodiment of the present application;

[0058] Figure 2 It is a first schematic diagram of a crack detection result provided by an embodiment of the present application;

[0059] Figure 3 It is a second schematic diagram of a crack detection result provided by an embodiment of the present application;

[0060] Figure 4 It is a third schematic diagram of a crack detection result provided by an embodiment of the present application;

[0061] Figure 5 It is a flowchart of an image detection method provided by an embodiment of the present application;

[0062] Figure 6 It is a flowchart of a training method for an initial object detection model provided by an embodiment of the present application;

[0063] Figure 7 It is a flowchart of a method for obtaining a sample road image provided by an embodiment of the present application;

[0064] Figure 8 It is a schematic diagram of a sample road image provided by an embodiment of the present application;

[0065] Figure 9 It is a schematic diagram of a global annotation box provided by an embodiment of the present application;

[0066] Figure 10 It is a schematic diagram of a local annotation box provided by an embodiment of the present application;

[0067] Figure 11 It is a flowchart of an incremental training method for an initial object detection model provided by an embodiment of the present application;

[0068] Figure 12 It is a schematic diagram of the architecture of a road crack detection device provided by an embodiment of the present application. Detailed Embodiments

[0069] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts belong to the scope of protection of the present application.

[0070] In the present application, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including one..." does not exclude the presence of additional identical elements in the process, method, article or device including the said element.

[0071] As Figure 1 shown, it is a flowchart of a road crack detection method provided by an embodiment of the present application, including the following steps:

[0072] S101: Obtain the road image to be detected.

[0073] S102: Input the road image to be detected into the target detection model to obtain the crack detection result.

[0074] Among them, the crack detection result at least includes the crack position (i.e., position coordinates) and crack type in the road image to be detected; the target detection model is obtained by incrementally training the initial target detection model using the cosine annealing algorithm; the initial target detection model is pre-trained based on the sample road images.

[0075] Optionally, the crack type includes but is not limited to longitudinal cracks and transverse cracks. Specifically, for the crack detection result, reference can be made to Figure 2 、 Figure 3 and Figure 4 . The red lines in the figure are cracks, and the crack position is represented in the form of [x1, y1, x2, y2] as the crack bounding box.

[0076] In step S102, the target detection model includes but is not limited to the DETR model.

[0077] Optionally, in another embodiment of the present application, the specific implementation manner of step S102, as Figure 5 shown, includes the following steps:

[0078] S501: Obtain the query vector in the target detection model.

[0079] Among them, the number of query vectors is usually fixed, and each query vector in the object detection model corresponds to a possible object (such as a crack).

[0080] S502: Input the road image to be detected into the encoder in the object detection model to obtain the road image features.

[0081] Among them, the road image features indicate the global features of the image.

[0082] It can be understood that the role of the encoder is to extract global features from the road image to be detected. Specifically, the object detection model first extracts local features of the image through a convolutional neural network (usually ResNet or other CNN architectures). The local features are input into the encoder, and the encoder transforms these features into higher-level global features, thereby capturing the relationships between various parts of the image and providing a global context perspective.

[0083] S503: Input the road image features and query vectors into the decoder in the object detection model to obtain the crack detection result.

[0084] It can be understood that the query vectors are used to guide the decoder to focus on different regions or objects in the image. The decoder will predict the position of each object in the image and the category of the object according to the query vectors and the global features provided by the encoder, thereby obtaining the crack detection result.

[0085] Optionally, in another embodiment of the present application, the specific implementation manner of pre-training the initial object detection model based on the sample road images, as Figure 6 shown, includes the following steps:

[0086] S601: Obtain sample road images.

[0087] Among them, the sample road images include road images under different road conditions and different lighting conditions, and various types and forms of road disease cracks are covered in the images. For example, 20,000 publicly available sample road images.

[0088] Optionally, in another embodiment of the present application, the specific implementation manner of step S601, as Figure 7 shown, includes the following steps:

[0089] S701: Obtain the initial road image.

[0090] Among them, the initial road image indicates the road image that has not undergone any processing.

[0091] S702: Perform normalization processing on the initial road image to obtain the processed initial road image.

[0092] It is understandable that the initial road image is standardized. Specifically, the initial road image is cropped according to a preset number of pixels to obtain multiple pixel images. For each pixel image, the pixel image is normalized to obtain a normalized pixel image, and histogram equalization is used to process the normalized pixel image to obtain the processed initial road image.

[0093] Optionally, the preset number of pixels includes, but is not limited to, 600*600.

[0094] It should be noted that if the number of pixels of the initial road image is less than the preset number of pixels, the initial road image with insufficient preset number of pixels is filled with black lines.

[0095] S703: Determine the processed initial road image as the sample road image.

[0096] Specifically, for the specific manifestation form of the sample road image, reference can be made to Figure 8 . Figure 8 It contains a total of 24 small sample road images, namely 1_10.jpg, 1_15.jpg, 2_5.jpg, 2_7.jpg, 2_8.jpg, 2_9.jpg, 2_10.jpg, 2_12.jpg, 2_15.jpg, 3_10.jpg, 4_8.jpg, 4_10.jpg, 6_7.jpg, 6_8.jpg, 6_12.jpg, 9_7.jpg, 9_8.jpg, 9_9.jpg, 9_10.jpg, 10_8.jpg, 11_7.jpg, 11_8.jpg, 12_12.jpg, and 12_13.jpg.

[0097] S602: Input the sample road image into the initial target detection model to obtain the sample crack detection result.

[0098] Among them, the sample crack detection result indicates the crack detection result corresponding to the sample road image.

[0099] S603: Calculate the loss function between the sample crack detection result and the actual crack detection result corresponding to the sample road image.

[0100] Among them, the loss function includes at least the cross-entropy loss function; the actual crack detection result contains the crack composite label; the crack composite label is composed of a local annotation box and a global annotation box; the global annotation box indicates the box containing the entire length of the crack; the local annotation box contains the box with a partial length of the crack.

[0101] It can be understood that by decomposing the cracks in the sample road image into N consecutive local annotation boxes and combining them with the global annotation box of the crack, a crack composite label is formed, achieving an accuracy level close to semantic segmentation while maintaining the compatibility of the object detection framework.

[0102] Specifically, for the specific manifestation form of the global annotation box of the sample road image, reference can be made to Figure 9 , Figure 9 where the red box in it is the global annotation box. For the specific manifestation form of the local annotation box of the sample road image, reference can be made to Figure 10 , Figure 10 and the green box in it is the local annotation box.

[0103] It can be seen from the above that through the adaptive local annotation box size determination mechanism, not only the crack length is considered, but also the change in crack width is combined. For the part of the crack with a large width change, the size of the local annotation box will be appropriately increased to include sufficient crack detail information and avoid missing key features due to too small an annotation box.

[0104] S604: When the loss function does not converge, use the Hungarian matching algorithm to adjust the model parameters of the initial object detection model and return to execute step S602.

[0105] It can be understood that when the loss function does not converge well, the threshold or cost function in the Hungarian matching algorithm can be dynamically adjusted, gradually increasing the matching tolerance or changing the matching strategy, so as to optimize the model parameters of the initial object detection model until the loss function of the model begins to decrease significantly.

[0106] S605: When the loss function converges, it is determined that the training of the initial object detection model is completed.

[0107] Among them, the initial object detection model is trained for 100 epochs, the batch_size is set to 16, the initial learning rate is 0.001, and the training is carried out using an RTX4080 24G graphics card. On the sample road image, the detection accuracy of the model for road targets reaches 88.4%. However, the generalization ability of the model on the new dataset is weak, and the accuracy is only 46.3%. Nevertheless, the model still has a strong ability to extract road crack features, laying a foundation for subsequent detection tasks.

[0108] Optionally, in another embodiment of the present application, the specific implementation manner of using the cosine annealing algorithm to incrementally train the initial object detection model to obtain the object detection model is as Figure 11 shown, including the following steps:

[0109] S1101: Obtain incremental sample road images.

[0110] Among them, the incremental sample road images indicate the road images collected by oneself. For example, 6,000 incremental sample road images.

[0111] S1102: Freeze the preset number of layers of the backbone network in the target detection model.

[0112] Among them, the backbone network includes but is not limited to ResNet, and the preset number of layers includes but is not limited to the first two layers.

[0113] It can be understood that in order to reduce the computational load, the first two layers of ResNet are frozen, and only the last three layers are retained, reducing redundant calculations while maintaining the detection ability for small targets.

[0114] S1103: After the freezing of the preset number of layers is completed, input the incremental sample road images into the target detection model to obtain incremental crack detection results.

[0115] Among them, the incremental crack detection results indicate the crack detection results corresponding to the incremental sample road images.

[0116] S1104: Calculate the loss function between the incremental crack detection results and the actual crack detection results corresponding to the incremental sample road images.

[0117] Optionally, the mean square error or cross-entropy loss function between the incremental crack detection results and the actual crack detection results corresponding to the incremental sample road images can be calculated and used as the loss function.

[0118] S1105: When the loss function does not converge, use the cosine annealing algorithm to adjust the learning rate, and return to execute step S1103 of inputting the incremental sample road images into the target detection model to obtain incremental crack detection results.

[0119] Among them, during the training of the base model, the initial learning rate is set to 1e-4, the minimum learning rate is 1e-6, and the learning rate decays by 10% every 10 epochs. When the loss function does not converge, the cosine annealing algorithm will be used to adjust the learning rate. During incremental training, the initial learning rate is also 1e-4, the minimum learning rate is 1e-6, and the learning rate decays by 10% every 6 epochs.

[0120] It can be understood that using the cosine annealing algorithm to adjust the learning rate enables the model to quickly adapt to the crack feature distribution in its own dataset. During the incremental training process, based on the general features learned in pre-training, the model further learns the features and detailed information of road disease cracks. By continuously adjusting the model parameters, the model adapts to the characteristics of a specific dataset, thereby improving the detection ability for road disease cracks.

[0121] S1106: When the loss function converges, determine the initial target detection model as the target detection model.

[0122] In summary, through the methods of pre-training and incremental training, the model can be adapted to the road disease crack detection tasks under different environments and conditions, which not only improves the detection efficiency but also ensures the accurate identification of road cracks.

[0123] As Figure 12 shown, it is a schematic architecture diagram of a road crack detection device provided by an embodiment of the present application. The detection device includes: an acquisition unit 100 and a detection unit 200.

[0124] The acquisition unit 100 is used to acquire the road image to be detected.

[0125] The detection unit 200 is used to input the road image to be detected into the target detection model to obtain the crack detection result; the crack detection result at least includes the crack position and crack type in the road image to be detected; the target detection model is obtained by incrementally training the initial target detection model using the cosine annealing algorithm; the initial target detection model is pre-trained based on the sample road images.

[0126] Specifically, the detection unit 200 is used to: obtain the query vector in the target detection model; input the road image to be detected into the encoder in the target detection model to obtain the road image feature; input the road image feature and the query vector into the decoder in the target detection model to obtain the crack detection result.

[0127] Specifically, the detection unit 200 is used to: obtain the incremental sample road image; freeze the preset number of layers of the backbone network in the target detection model; after the preset number of layers is frozen, input the incremental sample road image into the target detection model to obtain the incremental crack detection result; calculate the loss function between the incremental crack detection result and the actual crack detection result corresponding to the incremental sample road image; when the loss function does not converge, adjust the learning rate using the cosine annealing algorithm, and return to execute the step of inputting the incremental sample road image into the target detection model to obtain the incremental crack detection result; when the loss function converges, determine the initial target detection model as the target detection model.

[0128] In summary, through the methods of pre-training and incremental training, the model can be adapted to the road disease crack detection tasks under different environments and conditions, which not only improves the detection efficiency but also ensures the accurate identification of road cracks.

[0129] Combined with Figure 12 the content shown, the detection unit 200 includes:

[0130] The image acquisition unit is used to acquire the sample road image.

[0131] The image acquisition unit is specifically configured to: acquire an initial road image; perform normalization processing on the initial road image to obtain a processed initial road image; and determine the processed initial road image as a sample road image.

[0132] The input unit is configured to input the sample road image into the initial target detection model to obtain a sample crack detection result.

[0133] The calculation unit is configured to calculate a loss function between the sample crack detection result and the actual crack detection result corresponding to the sample road image; the loss function includes at least a cross-entropy loss function; the actual crack detection result includes a crack composite label; the crack composite label is composed of a local annotation box and a global annotation box; the global annotation box indicates a box containing the entire length of the crack; the local annotation box contains a box with a partial length of the crack.

[0134] The return unit is configured to, when the loss function does not converge, adjust the model parameters of the initial target detection model using the Hungarian matching algorithm and return to execute the step of inputting the sample road image into the initial target detection model to obtain a sample crack detection result.

[0135] The determination unit is configured to determine that the training of the initial target detection model is completed when the loss function converges.

[0136] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for a system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiment. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative work.

[0137] Those skilled in the art may further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.

[0138] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting road cracks, characterized in that, Including: Obtain the road image to be detected; Input the road image to be detected into the target detection model to obtain the crack detection result; The crack detection result at least includes the crack position and crack type in the road image to be detected; The target detection model is obtained by incrementally training the initial target detection model using the cosine annealing algorithm; The initial target detection model is pre-trained based on sample road images.

2. The method according to claim 1, wherein The step of inputting the road image to be detected into the target detection model to obtain the crack detection result includes: Obtain the query vector in the target detection model; Input the road image to be detected into the encoder in the target detection model to obtain the road image feature; Input the road image feature and the query vector into the decoder in the target detection model to obtain the crack detection result.

3. The method according to claim 1, characterized in that The process of pre-training the initial target detection model based on sample road images includes: Obtain sample road images; Input the sample road images into the initial target detection model to obtain sample crack detection results; Calculate the loss function between the sample crack detection result and the actual crack detection result corresponding to the sample road image; the loss function at least includes the cross-entropy loss function; the actual crack detection result contains a crack composite label; the crack composite label is composed of a local annotation box and a global annotation box; the global annotation box indicates a box containing the entire length of the crack; the local annotation box contains a box with a partial length of the crack; When the loss function does not converge, use the Hungarian matching algorithm to adjust the model parameters of the initial target detection model, and return to execute the step of inputting the sample road image into the initial target detection model to obtain the sample crack detection result; When the loss function converges, determine that the initial target detection model training is completed.

4. The method according to claim 3, wherein The step of obtaining sample road images includes: Obtain the initial road image; Perform normalization processing on the initial road image to obtain the processed initial road image; Determine the processed initial road image as the sample road image.

5. The method according to claim 1, wherein The process of incrementally training the initial target detection model using the cosine annealing algorithm to obtain the target detection model includes: Obtain incremental sample road images; Freeze a preset number of layers of the backbone network in the target detection model; After the preset number of layers are frozen, input the incremental sample road images into the target detection model to obtain incremental crack detection results; Calculate the loss function between the incremental crack detection result and the actual crack detection result corresponding to the incremental sample road image; When the loss function does not converge, use the cosine annealing algorithm to adjust the learning rate, and return to execute the step of inputting the incremental sample road images into the target detection model to obtain the incremental crack detection results; When the loss function converges, determine the initial target detection model as the target detection model.

6. A road crack detection device, characterized in that, Including: An acquisition unit for obtaining the road image to be detected; A detection unit for inputting the road image to be detected into the target detection model to obtain the crack detection result; The crack detection result at least includes the crack position and crack type in the road image to be detected; The target detection model is obtained by incrementally training the initial target detection model using the cosine annealing algorithm; The initial target detection model is pre-trained based on sample road images.

7. The device according to claim 6, characterized in that, The detection unit is specifically configured to: Obtain the query vector in the target detection model; Input the road image to be detected into the encoder in the target detection model to obtain road image features; Input the road image features and the query vector into the decoder in the target detection model to obtain the crack detection result.

8. The device according to claim 6, characterized in that The detection unit includes: An image acquisition unit for acquiring sample road images; An input unit for inputting the sample road image into the initial target detection model to obtain a sample crack detection result; A calculation unit for calculating the loss function between the sample crack detection result and the actual crack detection result corresponding to the sample road image; the loss function at least includes the cross-entropy loss function; the actual crack detection result includes a crack composite label; the crack composite label is composed of a local annotation box and a global annotation box; the global annotation box indicates a box containing the entire length of the crack; the local annotation box contains a box with a partial length of the crack; A return unit for, when the loss function does not converge, adjusting the model parameters of the initial target detection model using the Hungarian matching algorithm and returning to execute the step of inputting the sample road image into the initial target detection model to obtain a sample crack detection result; A determination unit for, when the loss function converges, determining that the training of the initial target detection model is completed.

9. The device according to claim 8, characterized in that, The image acquisition unit is specifically configured to: Obtain an initial road image; Perform normalization processing on the initial road image to obtain a processed initial road image; Determine the processed initial road image as the sample road image.

10. The device according to claim 6, characterized in that, The detection unit is specifically configured to: Obtain an incremental sample road image; Freeze a preset number of layers of the backbone network in the target detection model; After the preset number of layers is frozen, input the incremental sample road image into the target detection model to obtain an incremental crack detection result; Calculate the loss function between the incremental crack detection result and the actual crack detection result corresponding to the incremental sample road image; When the loss function does not converge, adjust the learning rate using the cosine annealing algorithm and return to execute the step of inputting the incremental sample road image into the target detection model to obtain an incremental crack detection result; When the loss function converges, determine the initial target detection model as the target detection model.

Images