CN-122024075-A - Capital construction apparent disease detection method, system and terminal based on AI large model

CN122024075ACN 122024075 ACN122024075 ACN 122024075ACN-122024075-A

Abstract

The invention discloses a method, a system and a terminal for detecting a basic construction apparent disease based on an AI large model, wherein the method comprises the steps of carrying out image enhancement processing on an image to be detected to obtain optimized detection data, constructing an improved SAM disease detection model, namely the AI large model, classifying pixels in the image to be detected by using the trained improved SAM disease detection model to obtain a classification result of the pixels, calculating the area of the disease according to the classified disease area, and positioning the specific position of the disease according to GPS information of the basic construction picture. According to the invention, the feature enhancement pretreatment is carried out on the foundation image to obtain a clearer detection sample with outstanding details, and the improved SAM network structure is combined to accurately detect the disease, so that a detector can detect the disease without approaching a dangerous area of the foundation, the potential safety hazard is eliminated, and the high-precision, rapid and automatic detection of the surface disease of the foundation is realized.

Inventors

HUANG HUI
LIU YANG
XIE KE
CAI QINGJUN
CHEN XIN

Assignees

深圳大学
中优云视(深圳)科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The method for detecting the apparent disease of the foundation based on the AI large model is characterized by comprising the following steps of: Constructing an improved SAM model comprising an improved hint encoder, an improved image encoder, and a mask decoder; Obtaining basic construction surface disease detection data, constructing a data set according to the basic construction surface disease detection data, and training the improved SAM model by using the data set to obtain a target disease detection model; shooting a foundation surface image to be predicted, and recording GPS coordinate information of a shooting position and geometric pose information of a shooting camera; preprocessing the foundation surface image to obtain a target foundation surface image, inputting the target foundation surface image into the target disease detection model for disease detection, and obtaining a segmentation mask of an area where the disease is located; and calculating the area of the disease according to the segmentation mask, and converting the image coordinates of the segmentation mask according to the GPS coordinate information and the geometric pose information to obtain the real world coordinates of the disease.
2. The AI-large-model-based construction apparent disease detection method of claim 1, wherein the improved hint encoder adds an adapter module based on an attention module of an original image encoder; The adapter module comprises a first adapter and a second adapter, wherein the first adapter is positioned at the downstream of the multi-head attention module of the attention module and before residual connection and is used for maintaining main information flow and optimizing the output of the multi-head attention module; the second adapter is located within the full connection layer residual path of the attention module, and introduces additional nonlinear transformations for enhancing learning capabilities of the model.
3. The AI-large-model-based capital construction apparent disease detection method according to claim 2, wherein the first adapter and the second adapter are both bottleneck structures; The first adapter and the second adapter are formed by mutually connecting a lower projection layer, a ReLU activation function and an upper projection layer in series; the formalized representation of the first adapter and the second adapter is: ; Wherein, the The input characteristics are represented as such, Representing the function of the ReLU activation, The lower projection layer is shown as such, The upper projection layer is shown as such, Representing the output result of the adapter.
4. The AI-large-model-based construction apparent disease detection method of claim 1, wherein the modified hint encoder selects a network CRSN for a disease area; The disease area selection network CRSN is used for automatically detecting and extracting information of potential disease areas in the image, generating sparse prompt codes and dense prompt codes according to the information, and inputting the sparse prompt codes and the dense prompt codes into a mask decoder.
5. The AI-large-model-based construction apparent disease detection method of claim 4, wherein the disease region selection network CRSN is a dual-flow architecture including dense hint branches and sparse hint branches; the dense prompting branch comprises a convolution layer, a batch normalization layer and an activation function, and the sparse prompting branch comprises the convolution layer, the batch normalization layer, the activation function, a maximum pooling layer and a full connection layer; The dense prompting branch and the sparse prompting branch both introduce an attention mechanism, and the dense prompting branch is used for adaptively fusing a prediction mask and a feature map through the attention mechanism to generate a first fusion feature and extracting first multi-modal information based on the first fusion feature; The sparse prompt branch is used for adaptively fusing the prediction mask and the feature map through an attention mechanism to generate a second fusion feature, and extracting second multi-modal information based on the second fusion feature; Wherein the first multi-modal information and the second multi-modal information are inputs to the mask decoder.
6. The AI-large-model-based construction apparent disease detection method according to claim 1, wherein the preprocessing the construction surface image to obtain a target construction surface image specifically comprises: performing histogram equalization and gamma correction on the brightness and contrast of the foundation surface image to obtain a radiation corrected image; Normalizing the color of the radiation corrected image through white balance correction and color space conversion to obtain a color normalized image; And carrying out geometric correction on the image after color standardization through homography transformation of MAGSAC ++ algorithm to obtain a target infrastructure surface image after geometric correction.
7. The AI-large-model-based construction apparent disease detection method according to claim 1, wherein the calculating the area of the disease according to the segmentation mask and converting the image coordinates of the segmentation mask according to the GPS coordinate information and the geometric pose information to obtain the real world coordinates of the disease specifically comprises: Calculating the disease area according to the number of pixels in the segmentation mask : ; Wherein, the Representing the number of pixels in the segmentation mask, Representing spatial resolution; and converting the coordinates of the segmentation mask from an image coordinate system to a real world coordinate system through inverse mapping of homography transformation according to the GPS coordinate information and the geometric pose information to obtain the real world coordinates of the diseases.
8. The utility model provides a capital construction apparent disease detecting system based on AI large model which characterized in that, the capital construction apparent disease detecting system based on AI large model includes: A model building module for building a modified SAM model comprising a modified hint encoder, a modified image encoder, and a mask decoder; The model training module is used for acquiring the detection data of the basic building surface diseases, constructing a data set according to the detection data of the basic building surface diseases, and training the improved SAM model by using the data set to obtain a target disease detection model; The image acquisition module is used for shooting a foundation surface image to be predicted and recording GPS coordinate information of a shooting position and geometric pose information of a shooting camera; the disease detection module is used for preprocessing the foundation surface image to obtain a target foundation surface image, inputting the target foundation surface image into the target disease detection model for disease detection, and obtaining a segmentation mask of an area where the disease is located; and the quantization positioning module is used for calculating the area of the disease according to the segmentation mask, and converting the image coordinates of the segmentation mask according to the GPS coordinate information and the geometric pose information to obtain the real world coordinates of the disease.
9. A terminal comprising a memory, a processor, and an AI-large-model-based constructed apparent disease detection program stored on the memory and operable on the processor, the AI-large-model-based constructed apparent disease detection program, when executed by the processor, implementing the AI-large-model-based constructed apparent disease detection method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing an AI-large model-based building apparent disease detection program which, when executed by a processor, implements the AI-large model-based building apparent disease detection method according to any one of claims 1 to 7.

Description

Capital construction apparent disease detection method, system and terminal based on AI large model Technical Field The invention relates to the technical field of infrastructure disease identification, in particular to a method, a system, a terminal and a computer-readable storage medium for detecting an apparent infrastructure disease based on an AI large model. Background The formation of foundation diseases is a complex process, and is mainly influenced by environmental factors (such as natural conditions of temperature change, precipitation, weathering and the like), structural stress (such as hydrostatic pressure, hydrodynamic pressure and foundation deformation born by a foundation), material degradation (such as aging, carbonization and chloride ion erosion of concrete materials) and the like. Once the disease is formed, aggressive medium in the external environment permeates into the interior of the foundation through the disease to cause continuous degradation of the concrete matrix, meanwhile, the disease provides a channel for water permeation, seepage pressure is increased, dangerous cases such as piping and soil flowing are possibly caused, the integrity of the foundation is weakened by expansion of the disease, the capability of the foundation for resisting external load is reduced, and finally the instability of the structure of the foundation is possibly caused. However, at present, the detection of the disease of the foundation is mainly dependent on professional technicians to adopt a disease depth finder and a disease width finder for regular inspection, but because the slope of the foundation is often provided with a large inclination, the traditional manual detection method has potential safety hazards, is influenced by human factors, environmental conditions and the precision of detection equipment, and is easy to miss detection and misjudge. In addition, the manual detection is long in time consumption, the large-area and high-frequency omnibearing monitoring is difficult to realize, the intermittent detection mode is also difficult to discover the dynamic development process of the disease in time, and the safety, accuracy and efficiency of the detection of the infrastructure disease are required to be improved. Accordingly, the prior art is still in need of improvement and development. Disclosure of Invention The invention mainly aims to provide a method, a system, a terminal and a computer readable storage medium for detecting a capital construction apparent disease based on an AI large model, and aims to solve the problems that the detection of the capital construction disease is easily influenced by external environment, is dependent on manual detection, is difficult to realize large-area and high-frequency omnibearing monitoring and has lower safety, accuracy and efficiency in the detection of the capital construction disease in the prior art. In order to achieve the above object, the present invention provides a method for detecting a building apparent disease based on an AI large model, the method for detecting a building apparent disease based on an AI large model comprising the steps of: Constructing an improved SAM model comprising an improved hint encoder, an improved image encoder, and a mask decoder; Obtaining basic construction surface disease detection data, constructing a data set according to the basic construction surface disease detection data, and training the improved SAM model by using the data set to obtain a target disease detection model; shooting a foundation surface image to be predicted, and recording GPS coordinate information of a shooting position and geometric pose information of a shooting camera; preprocessing the foundation surface image to obtain a target foundation surface image, inputting the target foundation surface image into the target disease detection model for disease detection, and obtaining a segmentation mask of an area where the disease is located; and calculating the area of the disease according to the segmentation mask, and converting the image coordinates of the segmentation mask according to the GPS coordinate information and the geometric pose information to obtain the real world coordinates of the disease. Optionally, in the AI large model-based construction apparent disease detection method, the improved hint encoder adds an adapter module based on the attention module of the original image encoder; The adapter module comprises a first adapter and a second adapter, wherein the first adapter is positioned at the downstream of the multi-head attention module of the attention module and before residual connection and is used for maintaining main information flow and optimizing the output of the multi-head attention module; the second adapter is located within the full connection layer residual path of the attention module, and introduces additional nonlinear transformations for enhancing learning capabilities of the model. Option