Visual identification method and system based on artificial intelligence

Abstract

The invention relates to the technical field of visual identification, in particular to a visual identification method and system based on artificial intelligence. Comprising the following steps: S1, acquiring image data acquired by a camera, preprocessing the image data, extracting a defect area in the image data by using a panoramic segmentation technology, and acquiring environment data during image shooting according to the image defect area; and S2, constructing a visual adaptation model according to the environmental data, constructing a prediction model to supplement the visual adaptation model, performing visual simulation on a defect area in the image data based on the visual adaptation model, and determining a shooting adjustment angle of the camera according to a simulation result. According to the invention, through image acquisition and preprocessing, panoramic segmentation, environmental data acquisition and visual adaptation model construction, a defect area is accurately identified; combining the prediction model and closed-loop control, dynamically adjusting camera parameters, and optimizing a defect detection effect; the accuracy and stability of the system are ensured through real-time feedback and evaluation, and the automation level and adaptability are improved.

Description

Technical Field

[0001] The present invention relates to the field of visual recognition technology, and in particular to a visual recognition method and system based on artificial intelligence. Background Art

[0002] Visual recognition usually includes image acquisition, preprocessing, feature extraction, target detection, classification recognition, tracking and semantic understanding. The advantages of existing visual recognition methods include high accuracy, robustness, real-time, versatility, adaptability and large-scale data processing capabilities. Artificial intelligence-based visual recognition methods further enhance these advantages, specifically manifested in higher accuracy, stronger robustness, more efficient real-time processing, wider versatility, better adaptability and more powerful data processing capabilities.

[0003] Despite this, existing visual recognition methods still have some disadvantages, mainly in the limited ability to handle complex scenes and rare situations, reliance on large amounts of labeled data, high consumption of computing resources, poor model interpretability, and unstable performance under extreme conditions. AI-based methods have brought significant improvements in these areas, including: 1. Improve the ability to handle complex scenarios and rare situations: Through deep learning and large-scale data training, AI models are able to better understand and handle complex and rare scenarios.

[0004] 2. Reduce dependence on large amounts of labeled data: Techniques such as semi-supervised learning, unsupervised learning, and transfer learning reduce the need for large amounts of labeled data.

[0005] 3. Reduce computing resource consumption: Reduce computing resource consumption through model compression, quantization and dedicated hardware acceleration.

[0006] 4. Enhance model interpretability: Through explainable machine learning techniques and visualization tools, the transparency and interpretability of the model are improved.

[0007] 5. Improve stability under extreme conditions: Through reinforcement learning and multimodal fusion, the stability and robustness of the model under extreme conditions are improved.

[0008] Especially in terms of stability, the AI-based visual recognition method significantly improves the stability and robustness under complex and extreme conditions, making the visual recognition system more reliable and efficient in various application scenarios. Summary of the invention

[0009] In order to overcome the shortcoming of poor stability of visual recognition, the present invention provides a visual recognition method and system based on artificial intelligence.

[0010] The technical solution of the present invention is: a visual recognition method based on artificial intelligence, comprising the following steps: S1: Obtain image data collected by the camera and perform preprocessing, extract defect areas in the image data using panoramic segmentation technology, and obtain environmental data when the image was taken based on the image defect areas; S2: Build a visual adaptation model based on environmental data, and build a prediction model to supplement the visual adaptation model. Perform visual simulation on the defective area in the image data based on the visual adaptation model, and determine the shooting adjustment angle of the camera based on the simulation results. S3: Dynamically adjust the camera height and angle through real-time feedback and prediction models, and closed-loop optimization ensures the accuracy and stability of defect detection.

[0011] Preferably, the method of extracting defective areas in image data using panoramic segmentation technology and obtaining environmental data when the image was taken based on the image defective areas includes: first, dividing the image into N areas using semantic segmentation and assigning a category label to each area; second, identifying the category of each area using instance segmentation and distinguishing different instances in the same category to obtain independent defective areas; finally, dividing each independent defective area into an adjustable area and an unadjustable area; after the classification is completed, obtaining environmental data of the unadjustable area in the taken image, the environmental data including lighting conditions and shooting distance, and the shooting distance is the distance between the camera and the defective area.

[0012] Preferably, the visual adaptation model is constructed according to the environmental data, including: obtaining the height data of the camera and the shooting defect range data and classifying them; dividing the camera into a height adjustment area, a transition area, and a manual observation area according to the height, wherein the height adjustment area is an area close to the camera and is used to adjust the height of the camera; the transition area is an intermediate area between the height adjustment area and the manual observation area and is used to smoothly connect the two main areas and provide additional adjustment space; the manual observation area is an area far away from the camera and is used for manual further observation; According to the defect range captured by the camera, the observation area is divided into two parts: 1. The directly captured defect area: The defect image directly captured by the camera is called the first captured image; 2. The manual observation area: The area that is not within the direct shooting range of the camera but is crucial for manual further evaluation of the defect. The image of this area is called the second captured image; In the manual observation area, a visual adaptation model is constructed; Among them, the second captured image is obtained from the first captured image. The specific obtaining process is as follows: First, obtain the first captured image taken by the camera. Secondly, use the camera to capture the image that conforms to the area for further manual observation. Finally, extract the environmental data and the data at the current moment of the image that conforms to the area for further manual observation; At the same time, a Transformer model is constructed, and the model is used to obtain the image data of the manually observed defect area in the area that conforms to the further manual observation, in the direction of directly looking at the defect and the side view of the defect. Then, the image data of the manually observed defect area is used as the second captured image.

[0013] Preferably, the visual simulation of the defect area in the image data based on the visual adaptation model includes: First, obtain the adjacent images that are close in time or space to the first captured image. The area of the defect area in the adjacent images is smaller than that in the first captured image; Extract the second captured image from the adjacent images; Obtain the actual distance from the manually observed position in the second captured image to the defect position in the first captured image and the light intensity at the current moment; Then, obtain K best observation areas through the formula; Among them, the process of obtaining the adjacent images is as follows: First, obtain the data at the current moment and the spatial position data of the first captured image, and use the RNN model to obtain L image data after the current moment. Secondly, extract the image data with a defect degree smaller than that of the image captured at the current moment from the L images. If there is image data with a defect degree smaller than that of the image captured at the current moment among the L images, then capture L1 image data at the adjacent position of the first captured image by the camera, L1 < L. Otherwise, continuously capture the area that conforms to the further manual observation by the camera and extract the image data with a light intensity less than or equal to the maximum light tolerance of the human eye, and obtain the light intensity data and the shooting position data in the image data.

[0014] Preferably, the obtaining of K best observation areas through the formula includes: The formula calculation is as follows, Among them, is the measurement value of the best observation condition, is the distance from the manually observed position to the defect position in the first captured image, is the distance from the camera to the defect position in the first captured image; is the light intensity, is the light intensity for human eye visual adaptation, is a constant less than 1.

[0015] Preferably, determining the shooting adjustment angle of the camera according to the simulation results includes: obtaining measurement value data of the best observation conditions for K best observation regions, extracting M eligible best observation regions, and sorting the M regions according to the change of light intensity over time in ascending order of the measurement values of the best observation conditions. M < K, and then sorting them according to the distance from the artificial observation position to the defect position of the first captured image to obtain the best observation region sequence.

[0016] Preferably, simultaneously constructing a prediction model to supplement the visual adaptation model includes: obtaining measurement value data of the best observation conditions, using the Kalman model to predict the adjustment values of the best shooting position and angle of the camera; then obtaining the training set of measurement value data of the best observation conditions, constructing an LSTM model, inputting the training set data into the LSTM model to obtain a trained LSTM model, using the LSTM model to predict the measurement values of the best observation conditions after the camera adjusts the shooting position and angle to obtain the predicted values of the measurement values of the best observation conditions; simultaneously generating a new best observation region sequence according to the rule for generating the best observation region sequence; simultaneously generating a new best observation region sequence according to the rule for generating the best observation region sequence.

[0017] Preferably, obtaining the best observation region sequence includes: based on the best observation region sequence and the new best observation region sequence, extracting the number of artificial observation positions in the best observation region sequence and the new best observation region sequence, and calculating the height adjustment amount of the camera through the formula; the formula calculation is as follows, Where, is the height adjustment amount of the camera, is the continuity of the th best observation region sequence, is the continuity of the th new best observation region sequence. The continuity is the change of the distance of the artificial observation position, and the distance change is the change of the distance from the artificial observation position to the defect position of the first captured image. and are constants, and are respectively the number of artificial observation positions in the th best observation region sequence and the new best observation region sequence, is the total number of the best observation region sequences.

[0018] Preferably, the camera height and angle are dynamically adjusted through real-time feedback and prediction models, and closed-loop optimization ensures the accuracy and stability of defect detection, including: designing the system as a closed-loop control loop, continuously adjusting and optimizing the camera parameters through a real-time feedback mechanism; using a model-based optimization algorithm, combined with real-time feedback data, to dynamically adjust the camera height and angle.

[0019] A visual recognition system based on artificial intelligence, comprising: Image acquisition and preprocessing module: uses a high-resolution industrial camera to continuously capture images at a fixed frame rate and optimizes image quality through denoising, color correction, contrast enhancement, and size adjustment preprocessing steps; Panoramic segmentation and defect area extraction module: uses deep learning models for semantic segmentation and combines MaskR-CNN to achieve instance segmentation, accurately identifying and classifying defect areas in the image into adjustable areas and non-adjustable areas; Environmental data acquisition module: deploys multiple sensors to monitor the environmental conditions during shooting in real time, including light intensity, temperature and humidity, and develops software tools to evaluate the impact of these factors on image quality and assist in image interpretation and correction; Visual adaptation model building and simulation module: Use GPU to accelerate ray tracing and Transformer model, simulate the best observation conditions based on environmental data, and optimize the camera shooting angle; Shooting angle and height adjustment module: Integrates electric pan / tilt or linear actuator and gyroscope equipment to allow remote control of the camera's position and posture, automatically adjusts to the optimal settings by receiving results from the visual adaptation model, uses Kalman filtering and LSTM models to predict the camera's optimal shooting position and angle, and dynamically generates a new sequence of optimal observation areas; Evaluation and Verification Module: Establish test benchmarks to regularly evaluate system performance, develop anomaly detection algorithms to monitor operating status, and build a closed-loop control system to collect feedback information for continuous improvement.

[0020] Beneficial effects: First, the defect area is extracted through panoramic segmentation technology, ensuring the high quality of image data and high accuracy of defect detection. Secondly, by constructing a visual adaptation model and a prediction model, the robustness and accuracy of the model under different environmental conditions are improved, ensuring the accurate identification of the defect area. Then, through visual simulation and height adjustment, the shooting angle and height of the camera are dynamically optimized to ensure the best defect detection effect and improve the adaptability and flexibility of the system. Further, through a closed-loop control loop and a real-time feedback mechanism, the height and angle of the camera are dynamically adjusted to ensure the continuous optimization and stable operation of the system, thereby improving the automation level of the system. In addition, through the evaluation and verification mechanism, the accuracy and stability of defect detection are ensured, the interpretability and credibility of the system are improved, and the user's trust is enhanced. When the camera cannot identify or process the defect area, through visual simulation and height adjustment, the area that meets the human observation is found, ensuring that people at the position of human observation can observe the defect area captured by the camera on the spot, thereby improving the reliability and practicality of the system. Finally, the height adjustment amount of the camera is calculated through a formula, which not only optimizes the height of the camera, but also utilizes the situation of the manual observation position to further improve the adjustment of the camera, thereby achieving the purpose of reducing the number of manual observations. At the same time, it provides the best position for manual observation and improves the overall performance and efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of a visual recognition method and system based on artificial intelligence of the present invention; Figure 2 The figure is a schematic diagram of the structure of a visual recognition system based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] Embodiment 1: A visual recognition method based on artificial intelligence, such as Figure 1 As shown, the following steps are included: S1: Obtain image data collected by the camera and perform preprocessing, extract defect areas in the image data using panoramic segmentation technology, and obtain environmental data when the image was taken based on the image defect areas; S2: Build a visual adaptation model based on environmental data, and build a prediction model to supplement the visual adaptation model. Perform visual simulation on the defective area in the image data based on the visual adaptation model, and determine the shooting adjustment angle of the camera based on the simulation results. S3: Dynamically adjust the camera height and angle through real-time feedback and prediction models, and closed-loop optimization ensures the accuracy and stability of defect detection.

[0024] The panorama segmentation technology is used to extract defect areas in image data, and the environmental data when the image is shot is obtained according to the image defect areas, including: first, the image is divided into N areas by semantic segmentation, and a category label is assigned to each area; secondly, the category of each area is identified by instance segmentation, and different instances in the same category are distinguished to obtain independent defect areas; finally, each independent defect area is divided into an adjustable area and an unadjustable area; after the classification is completed, the environmental data of the unadjustable area in the shot image is obtained, and the environmental data includes lighting conditions and shooting distance, and the shooting distance is the distance between the camera and the defect area.

[0025] Further explanation is that the defect area includes an area that is not clearly photographed or an area where the computer cannot identify defects; the adjustable area is a defect area improved by existing technology, and the non-adjustable area refers to a defect area that cannot be processed by existing technology; the panoramic segmentation technology is used to extract the defect area in the image data, and the environmental data during shooting is obtained, so that the defects can be accurately identified and located, while considering the lighting and distance environmental factors, so as to improve the accuracy and robustness of detection; the specific steps include semantic segmentation and instance segmentation to ensure that each defect area is independently identified and classified, thereby optimizing subsequent adjustments and processing; the independent defect area refers to different defect instances distinguished in the same category.

[0026] A visual adaptation model is constructed based on environmental data, including: obtaining the camera's height data and the defect range data and classifying them; dividing the camera into a height adjustment area, a transition area, and a manual observation area according to the height. The height adjustment area is an area close to the camera and is used to adjust the camera's height; the transition area is an intermediate area between the height adjustment area and the manual observation area, which is used to smoothly connect the two main areas and provide additional adjustment space; the manual observation area is an area far away from the camera and is used for further manual observation; according to the defect range captured by the camera, the observation area is divided into two parts: 1. Directly captured defect area: the defect image directly captured by the camera is called the first captured image; 2. Manual observation area: the defect area not directly captured by the camera The image of the area that is within the range but is crucial for manual further evaluation of defects is called the second captured image; in the manual observation area, a visual adaptation model is constructed; wherein the second captured image is obtained through the first captured image, and the specific acquisition process is firstly to obtain the first captured image taken by the camera, and then to use the camera to capture an image that meets the requirements of the manual further observation area, and finally to extract the environmental data and current time data that meet the requirements of the image of the manual further observation area; at the same time, a Transformer model is constructed, and the model is used to obtain the image data of the manual observation defect area in the manual direct defect direction and the side defect direction in the manual further observation area, and then the image data of the manual observation defect area is used as the second captured image.

[0027] Further explanation is that the purpose of obtaining the height of the camera is to adjust the height of the camera to minimize the degree of defects in the image taken by the camera by analyzing the degree of defects in the image taken by the camera at the current height; at the same time, the purpose of obtaining the height of the camera is to divide the height of the camera into a height adjustment area, a transition area, and a manual observation area, so as to achieve the purpose of identifying unclear areas through manual field observation if the camera cannot shoot clearly or recognize inadequately; the purpose of dividing the camera into a shooting defect area and a manual observation area corresponding to the defect area according to the shooting defect range is to find out the area that can be manually observed by obtaining the shooting range of the camera, and obtain the best manual observation position and observation angle from the area that can be manually observed; the purpose of constructing a visual adaptation model is to quantify the manually observable position and observation angle by constructing a visual adaptation model, because vision is affected by observation distance and light intensity, and the observation effect is different under the influence of different observation distances and light intensities.

[0028] Visual simulation of defect areas in image data based on a visual adaptation model, including: First, obtain adjacent images that are close in time or space to the first captured image, where the area of the defect area in the adjacent image is smaller than that in the first captured image; extract the second captured image from the adjacent images; obtain the actual distance from the artificial observation position in the second captured image to the defect position in the first captured image and the light intensity at the current moment; then obtain K best observation areas through a formula; Among them, the process of obtaining adjacent images is as follows: First, obtain the data of the current moment and the spatial position data of the first captured image, use the RNN model to obtain L image data after the current moment, and then extract the image data with a defect degree less than that of the captured image at the current moment from the L images. If there is image data with a defect degree less than that of the captured image at the current moment among the L images, then capture L1 image data at the adjacent position of the first captured image by the camera, where L1 < L. Otherwise, continuously capture the images that meet the artificial further observation area by the camera and extract the image data with a light intensity less than or equal to the maximum light tolerance of the human eye to obtain the light intensity data and the shooting position data in the image data.

[0029] Obtain K best observation areas through a formula, including: The formula calculation is as follows, Among them, is a measure value of the best observation condition, is the distance from the artificial observation position to the defect position in the first captured image, is the distance from the camera to the defect position in the first captured image; is the light intensity, is the light intensity of human eye visual adaptation, is a constant less than 1.

[0030] For further explanation, the explanations of each part are as follows: Indicates that the area is more suitable as the best observation area; Indicates the distance from vision to the observation area; The visual observation distance has a direct impact on the observation effect; The closer the distance, the clearer the observation details, but the field of view is limited; The farther the distance, the wider the field of view, but the details are blurred; Indicates the distance from the defect to the camera; The actual distance of the defect affects the visibility and clarity of the defect; The closer the distance, the easier it is to identify the defect; The farther the distance, the defect may become blurred; Indicates the light intensity of the current environment; The light intensity has a significant impact on the observation effect; The higher the light intensity, the brighter the image and the clearer the details; The lower the light intensity, the darker the image and the details are not easy to identify; Represents the adaptation light intensity of vision in the current environment; the adaptation light intensity of vision reflects the degree of adaptation of vision to the current environment; if the environmental light intensity differs greatly from the adaptation light intensity of vision, it will lead to poor observation effects; Used to adjust the growth rate of the exponential part; The value determines the growth rate of the exponential part; a small value will make the exponential part grow slowly, and a large value will make the exponential part grow rapidly, thus affecting the calculation of the value; For the specific explanation of the formula, the logarithmic part ( ) is used to calculate the difference between the visual observation distance and the actual distance of the defect; the exponential part ( ) is used to calculate the difference between the light intensity and the adaptation light intensity of vision; The value is used to adjust the growth rate of the exponential part to make it more flexible; by introducing the constant , the formula adjusts the growth rate of the exponential part according to the actual situation, making the model more flexible and adaptable.

[0031] Determine the shooting adjustment angle of the camera according to the simulation results, including: obtaining the measurement value data of the best observation conditions for K best observation areas, extracting M eligible best observation areas, and sorting the M areas according to the change of light intensity over time in ascending order of the measurement values of the best observation conditions. Since M < K, then sort them according to the distance from the artificial observation position to the defect position in the first captured image to obtain the best observation area sequence.

[0032] At the same time, construct a prediction model to supplement the visual adaptation model, including: obtaining the measurement value data of the best observation conditions, using the Kalman model to predict the adjustment values of the best shooting position and angle of the camera; then obtaining the training set of the measurement value data of the best observation conditions, constructing an LSTM model, inputting the training set data into the LSTM model to obtain the trained LSTM model, using the LSTM model to predict the measurement value of the best observation conditions after the camera adjusts the shooting position and angle to obtain the predicted value of the measurement value of the best observation conditions; at the same time, generate a new best observation area sequence according to the rule of generating the best observation area sequence.

[0033] For further explanation, construct an LSTM model to supplement the visual adaptation model, optimize the model through the training set data, predict the measurement value of the best observation conditions, and improve the accuracy and robustness of the model.

[0034] Obtaining the best observation area sequence includes: extracting the number of manual observation positions of the best observation area sequence and the new best observation area sequence according to the best observation area sequence and the new best observation area sequence, and calculating the height adjustment amount of the camera by a formula; the calculation formula is as follows, in, is the camera height adjustment, For the The continuity of the sequence of best observation areas, For the The continuity of a new optimal observation area sequence, wherein the continuity is the change of the distance of the manual observation position, and the distance change is the change of the distance from the manual observation position to the defect position of the first captured image, and is a constant, and Respectively The number of manual observation positions of the best observation area sequence and the new best observation area sequence, is the total number of the best observed region sequences.

[0035] Further explanation is that the continuity reflects the smooth change of the distance of the manual observation position. If the change is smooth, it means that the adjustment of the observation position is more reasonable and stable; the smoothness of the distance change reflects the rationality of the adjustment of the observation position. If the change is smooth, it means that the adjustment is more reasonable and stable; Explanation of each part: Indicates the camera height that needs to be adjusted. A positive value indicates an increase, while a negative value indicates a decrease. represents the number of optimal observation areas; Indicates that before adjustment, The initial state of the best observation area reflects the continuity of the best observation area before adjustment; It means that after adjustment, The expected state of the best observation area reflects the continuity of the best observation area after adjustment. and The value of adjusts the sensitivity and stability of the formula to prevent excessive fluctuations; and , the sensitivity and stability of the control formula; It indicates the number of manual observation positions expected to be required after the adjustment; it reflects the expected manual observation demand after the adjustment; Indicates the number of manual observation positions actually required before adjustment; reflects the actual manual observation demand before adjustment; Detailed explanation of the formula, exponential part: Calculation: , Function: Calculate the contribution of each optimal observation area to the height adjustment; the ratio reflects the change of the optimal observation area before and after the adjustment; if and The difference is not big, the ratio is close to 1, and the value of the exponential part is close to 1; if Much greater than , the ratio will increase, the value of the exponential part will decrease; the difference - : The difference reflects the change in the number of manual observations before and after the adjustment; if Less than , the difference is negative, and the exponential part will increase; if Greater than , the difference is positive, and the value of the exponential part will decrease; the height adjustment contributions of all the best observation areas are summed up to get the total adjustment amount; through the summation, the formula comprehensively considers the adjustment requirements of all the best observation areas, making the final height adjustment amount more reasonable and comprehensive; The formula comprehensively considers multiple factors such as the original state, predicted state, and changes in the number of manual observations of each optimal observation area to ensure that the calculation of the adjustment amount is more comprehensive and accurate; by introducing the constant and The formula adjusts the sensitivity and stability of the formula according to the actual situation, making it more flexible and adaptable; through comprehensive consideration of multiple factors, the formula can better adapt to different observation conditions and improve the robustness and stability of the system; by calculating the contribution of each optimal observation area to the height adjustment amount, the formula can optimize the height adjustment of the camera, making the defect detection effect more ideal; the formula supports dynamic adjustment, and adjusts according to real-time observation data and environmental conditions to ensure the performance of the system in different scenarios.

[0036] The camera height and angle are dynamically adjusted through real-time feedback and prediction models, and closed-loop optimization ensures the accuracy and stability of defect detection, including: designing the system as a closed-loop control loop to continuously adjust and optimize the camera parameters through a real-time feedback mechanism; using a model-based optimization algorithm combined with real-time feedback data to dynamically adjust the camera height and angle.

[0037] Embodiment 2: Based on Embodiment 1, a visual recognition system based on artificial intelligence includes: Image acquisition and preprocessing module: uses a high-resolution industrial camera to continuously capture images at a fixed frame rate and optimizes image quality through denoising, color correction, contrast enhancement, and size adjustment preprocessing steps; Panoramic segmentation and defect area extraction module: uses deep learning models for semantic segmentation and combines MaskR-CNN to achieve instance segmentation, accurately identifying and classifying defect areas in the image into adjustable areas and non-adjustable areas; Environmental data acquisition module: deploys multiple sensors to monitor the environmental conditions during shooting in real time, including light intensity, temperature and humidity, and develops software tools to evaluate the impact of these factors on image quality and assist in image interpretation and correction; Visual adaptation model building and simulation module: Use GPU to accelerate ray tracing and Transformer model, simulate the best observation conditions based on environmental data, and optimize the camera shooting angle; Shooting angle and height adjustment module: Integrates electric pan / tilt or linear actuator and gyroscope equipment to allow remote control of the camera's position and posture, automatically adjusts to the optimal settings by receiving results from the visual adaptation model, uses Kalman filtering and LSTM models to predict the camera's optimal shooting position and angle, and dynamically generates a new sequence of optimal observation areas; Evaluation and Verification Module: Establish test benchmarks to regularly evaluate system performance, develop anomaly detection algorithms to monitor operating status, and build a closed-loop control system to collect feedback information for continuous improvement.

[0038] The above is a detailed introduction to the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A visual recognition method based on artificial intelligence, characterized in that: The following steps are involved: S1: Obtain image data collected by the camera and perform preprocessing, extract defect areas in the image data using panoramic segmentation technology, and obtain environmental data when the image was taken based on the image defect areas; S2: Build a visual adaptation model based on environmental data, and build a prediction model to supplement the visual adaptation model. Perform visual simulation on the defective area in the image data based on the visual adaptation model, and determine the shooting adjustment angle of the camera based on the simulation results. S3: Dynamically adjust the camera height and angle through real-time feedback and prediction models, and closed-loop optimization ensures the accuracy and stability of defect detection.

2. The method for visual recognition based on artificial intelligence as claimed in claim 1, characterized in that: The method uses panoramic segmentation technology to extract defect areas in image data, and obtains environmental data when the image is shot according to the image defect areas, including: first, using semantic segmentation to divide the image into N areas, and assigning a category label to each area; second, using instance segmentation to identify the category of each area, and distinguishing different instances in the same category to obtain independent defect areas; finally, each independent defect area is divided into an adjustable area and an unadjustable area; after the classification is completed, the environmental data of the unadjustable area in the shot image is obtained, and the environmental data includes lighting conditions and shooting distance, and the shooting distance is the distance between the camera and the defect area.

3. The artificial intelligence-based visual recognition method according to claim 1, characterized in that: The visual adaptation model is constructed according to the environmental data, including: obtaining the height data of the camera and the shooting defect range data and classifying them; dividing the camera into a height adjustment area, a transition area, and a manual observation area according to the height, wherein the height adjustment area is an area close to the camera and is used to adjust the height of the camera; the transition area is an intermediate area between the height adjustment area and the manual observation area and is used to smoothly connect the two main areas and provide additional adjustment space; the manual observation area is an area far away from the camera and is used for manual further observation; According to the defect range captured by the camera, the observation area is divided into two parts:

1. Directly captured defect area: the defect image directly captured by the camera is called the first captured image; 2. Artificial observation area: the area that is not within the direct shooting range of the camera but is crucial for further manual evaluation of defects. The image of this area is called the second captured image. In the artificial observation area, a visual adaptation model is constructed; Among them, the second captured image is obtained through the first captured image, and the specific acquisition process is: first, the first captured image taken by the camera is obtained, and then the image of the area that conforms to the further manual observation is taken by the camera, and finally the environmental data and current time data that conform to the image of the area that is further manual observation are extracted; at the same time, a Transformer model is constructed, and the model is used to obtain the image data of the artificially observed defect area in the direction of the artificial direct defect and the direction of the side defect in the area that conforms to the further manual observation, and then the image data of the artificially observed defect area is used as the second captured image.

4. The method for visual recognition based on artificial intelligence as claimed in claim 1, characterized in that: The visual simulation of the defective area in the image data based on the visual adaptation model includes: First, obtain adjacent images that are close in time or space to the first captured image, and the area of the defective area in the adjacent images is smaller than that in the first captured image; extract the second captured image from the adjacent images; obtain the actual distance from the artificial observation position in the second captured image to the defective position in the first captured image and the light intensity at the current moment; then obtain K best observation areas through a formula; Among them, the process of obtaining adjacent images is as follows: First, obtain the data of the current moment and the spatial position data of the first captured image, use the RNN model to obtain L image data after the current moment, and then extract the image data with a defect degree less than that of the captured image at the current moment from the L images. If there is image data with a defect degree less than that of the captured image at the current moment among the L images, then capture L1 image data at the adjacent position of the first captured image by the camera, where L1 < L. Otherwise, continuously capture the images in the area that meets the artificial further observation area by the camera, and extract the image data with a light intensity less than or equal to the maximum light tolerance of the human eye, and obtain the light intensity data and the shooting position data in the image data.

5. The method for visual recognition based on artificial intelligence as claimed in claim 4, characterized in that: The formula for obtaining K best observation areas includes: the formula is calculated as follows: in, is a measure of the optimal observation conditions, is the distance from the manual observation position to the defect position of the first captured image, is the distance from the camera to the defect position in the first captured image; is the light intensity, To adapt the light intensity to human vision, is a constant less than 1.

6. The method for visual recognition based on artificial intelligence as claimed in claim 1, characterized in that: The determination of the shooting adjustment angle of the camera according to the simulation result includes: obtaining the measurement value data of the best observation conditions of the K best observation areas, extracting M eligible best observation areas, sorting the M areas according to the change of light intensity over time in ascending order of the measurement value of the best observation conditions, where M < K, and then sorting them according to the distance from the artificial observation position to the defective position in the first captured image to obtain the best observation area sequence.

7. The method for visual recognition based on artificial intelligence as claimed in claim 1, characterized in that: The simultaneous construction of a prediction model to supplement the visual adaptation model includes: obtaining the measurement value data of the best observation conditions, using the Kalman model to predict the adjustment values of the best shooting position and angle of the camera; then obtaining the training set of the measurement value data of the best observation conditions, constructing an LSTM model, inputting the training set data into the LSTM model to obtain a trained LSTM model, using the LSTM model to predict the measurement value of the best observation conditions after the camera adjusts the shooting position and angle, and obtaining the predicted value of the measurement value of the best observation conditions; at the same time, generating a new best observation area sequence according to the rule of generating the best observation area sequence.

8. The method for visual recognition based on artificial intelligence as claimed in claim 6, characterized in that: The obtaining of the optimal observation area sequence includes: extracting the number of manual observation positions of the optimal observation area sequence and the new optimal observation area sequence according to the optimal observation area sequence and the new optimal observation area sequence, and calculating the height adjustment amount of the camera by a formula; the calculation formula is as follows: in, is the camera height adjustment, For the The continuity of the sequence of best observation areas, For the The continuity of a new optimal observation area sequence, the continuity is the change of the distance of the manual observation position, the distance change is the change of the distance from the manual observation position to the defect position of the first captured image, and is a constant, and Respectively The number of manual observation positions of the best observation area sequence and the new best observation area sequence, is the total number of the best observed region sequences.

9. The method for visual recognition based on artificial intelligence as claimed in claim 1, characterized in that: The dynamic adjustment of the camera height and angle through real-time feedback and the prediction model, and the closed-loop optimization ensure the accuracy and stability of defect detection, including: designing the system as a closed-loop control loop, continuously adjusting and optimizing the parameters of the camera through a real-time feedback mechanism; adopting a model-based optimization algorithm, combining real-time feedback data, and dynamically adjusting the height and angle of the camera.

10. An artificial intelligence-based visual recognition system, used in the artificial intelligence-based visual recognition method according to any one of claims 1 to 9, characterized in that: It includes: Image acquisition and preprocessing module: Use a high-resolution industrial camera to continuously capture images at a fixed frame rate, and optimize the image quality through preprocessing steps such as denoising, color correction, contrast enhancement, and size adjustment; Panoramic segmentation and defect area extraction module: uses deep learning models for semantic segmentation and combines MaskR-CNN to achieve instance segmentation, accurately identifying and classifying defect areas in the image into adjustable areas and non-adjustable areas; Environmental data acquisition module: deploys multiple sensors to monitor the environmental conditions during shooting in real time, including light intensity, temperature and humidity, and develops software tools to evaluate the impact of these factors on image quality and assist in image interpretation and correction; Visual adaptation model building and simulation module: Use GPU to accelerate ray tracing and Transformer model, simulate the best observation conditions based on environmental data, and optimize the camera shooting angle; Shooting angle and height adjustment module: Integrates electric pan / tilt or linear actuator and gyroscope equipment to allow remote control of the camera's position and posture, automatically adjusts to the optimal settings by receiving results from the visual adaptation model, uses Kalman filtering and LSTM models to predict the camera's optimal shooting position and angle, and dynamically generates a new sequence of optimal observation areas; Evaluation and Verification Module: Establish test benchmarks to regularly evaluate system performance, develop anomaly detection algorithms to monitor operating status, and build a closed-loop control system to collect feedback information for continuous improvement.

Images