A deep neural network-based thermal infrared fixed-focus camera calibration method and system

Abstract

The application discloses a kind of based on deep neural network's thermal infrared fixed focus camera calibration method and system, the deep neural network based on thermal infrared fixed focus camera calibration method of the present application includes the following steps: the pixel coordinates of corner point in checkerboard picture are identified to checkerboard picture using pre-trained target detection model based on depth neural network;The pixel coordinates of all corner points are corrected as a whole using homography matrix to obtain corrected pixel coordinates;According to the pixel coordinates of each corner point after correction, camera calibration is carried out using Zhang Zhengyou camera calibration method to obtain the intrinsic matrix and distortion coefficient of camera.The present application aims at using the deep learning model that has learned widely can robustly extract the position of corner point in image, overcome the problem of imaging blur caused by low resolution, low signal-to-noise ratio and fixed focal length of infrared image, provide a kind of stable and reliable calibration method for commonly used fixed focus infrared camera.

Description

Technical Field

[0001] This invention relates to the field of camera calibration technology, specifically to a method and system for calibrating a thermal infrared fixed-focus camera based on a deep neural network. Background Technology

[0002] Infrared thermal imaging is an advanced imaging method that captures the invisible infrared radiation emitted by objects and transforms it into a visual image. The core of this technology lies in the direct correlation between an object's temperature and the energy of its emitted infrared radiation. All objects radiate infrared energy according to their own temperature, and infrared thermal imaging cameras can sensitively capture these minute radiations, convert them into electrical signals, and ultimately generate an image showing the temperature distribution on the object's surface. Although infrared imaging technology can provide rich temperature data, to accurately determine the specific temperature at each location, this data must be combined with the geometric information of the image. To obtain accurate geometric information, a geometric calibration process must be performed to determine the camera's intrinsic parameter matrix and distortion coefficients. This process ensures the accuracy and reliability of infrared images, allowing us to interpret and apply these images more accurately. However, infrared imaging typically has low resolution, contrast, and signal-to-noise ratio, which significantly affects the accurate detection of corner points in infrared images, thus impacting calibration accuracy. More importantly, most thermal infrared cameras are non-zoomable, inevitably leading to out-of-focus situations during calibration image acquisition. Therefore, the field of infrared camera calibration urgently needs a robust, reliable, and high-precision camera calibration method. Deep learning algorithms are now widely used in image processing, and numerous studies have shown that deep learning-based image processing methods have gradually become a stable and reliable technology. Camera calibration, as a method of extracting information from a large number of calibration object images to estimate camera parameters, shares inherent similarities with data-driven deep learning techniques. However, how to specifically implement thermal infrared fixed-focus camera calibration based on deep neural networks remains a key technical problem that urgently needs to be solved. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a calibration method and system for thermal infrared fixed-focus cameras based on deep neural networks, addressing the aforementioned problems in the prior art. This invention aims to utilize a deeply learned deep learning model to robustly extract corner positions in images, overcoming the problems of low resolution, low signal-to-noise ratio, and image blurring caused by fixed focal length in infrared images, and providing a stable and reliable calibration method for commonly used fixed-focus infrared cameras.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A calibration method for a thermal infrared fixed-focus camera based on a deep neural network includes the following steps: S1, using a pre-trained target detection model based on a deep neural network to identify the pixel coordinates of the corner points in a checkerboard image; S2, using a homography matrix to correct the pixel coordinates of all corner points to obtain the corrected pixel coordinates; S3, using the Zhang Zhengyou camera calibration method to calibrate the camera based on the corrected pixel coordinates of each corner point to obtain the camera's intrinsic parameter matrix and distortion coefficients.

[0006] Optionally, the target detection model is an improved version of the YOLOv8 target detection model. The improved version of the YOLOv8 target detection model adds a cascaded Local Window Attention (LWA) module to the output of the backbone network of the YOLOv8 target detection model to improve the network's ability to mine local information in the image.

[0007] Optionally, the Local Window Attention Mechanism (LWA) module processes the input features from the output of the backbone network by: first, dividing the input feature map into non-overlapping windows; then, sequentially stitching each window together on the depth channel; calculating the attention mechanism with the window center as the anchor point; and finally, returning the calculation result to the position of the anchor point in the original feature map.

[0008] Optionally, step S1 may be preceded by a step of training an object detection model using manually labeled checkerboard images.

[0009] Optionally, in step S2, the homography matrix is ​​used to perform a global correction on the coordinates of all corner points to obtain the corrected corner point coordinates, including:

[0010] S2.1, based on the pixel coordinates of the n corner points in the identified chessboard image. and their corresponding world coordinates Substituting into the equation shown below, we obtain a total of 2n equations:

[0011] ,

[0012] In the above formula, , , , , , , and Homography matrix The elements in the set, and have:

[0013] ;

[0014] S2.2, Solve the 2n equations using the least squares method to obtain the homography matrix. Elements in;

[0015] S2.3, homography matrix Substituting the elements in the formula into the overall formula, we can obtain the corrected corner coordinates by correcting all corner coordinates:

[0016] ,

[0017] In the above formula, These are the pixel coordinates after corner correction.

[0018] Optionally, in step S3, the camera intrinsic parameter matrix and distortion coefficients are obtained by using Zhang Zhengyou's camera calibration method based on the corrected pixel coordinates of each corner point. This involves solving the optimization problem shown in the following formula to finally obtain the camera's intrinsic parameter matrix and distortion coefficients:

[0019] ,

[0020] In the above formula, The number of chessboard images. This represents the number of corner points in each chessboard image. Let J be the corrected pixel coordinates of the j-th corner point in the i-th chessboard image. Let J be the coordinates of the reprojection point of the j-th corner point in the i-th chessboard image. This is the intrinsic parameter matrix of the camera. Let be the rotation matrix of the i-th chessboard image. Let be the translation vector of the i-th chessboard image. Let j be the world coordinates of the j-th corner point in the i-th chessboard image. These are the distortion coefficients, including radial distortion coefficients and tangential distortion coefficients.

[0021] Furthermore, the present invention also provides a thermal infrared fixed-focus camera calibration system based on a deep neural network, comprising:

[0022] The corner recognition program unit is used to identify the coordinates of corner points in a chessboard image using a pre-trained object detection model based on a deep neural network.

[0023] The corner coordinate correction program unit is used to perform a global correction on the coordinates of all corner points using the homography matrix to obtain the corrected corner coordinates.

[0024] The parameter calibration program unit is used to calibrate the camera using the Zhang Zhengyou camera calibration method based on the corrected corner coordinates to obtain the camera's intrinsic parameter matrix and distortion coefficients.

[0025] Furthermore, the present invention also provides a thermal infrared fixed-focus camera calibration system based on a deep neural network, comprising a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the thermal infrared fixed-focus camera calibration method based on a deep neural network.

[0026] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the deep neural network-based thermal infrared fixed-focus camera calibration method by a processor.

[0027] In addition, the present invention also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the deep neural network-based thermal infrared fixed-focus camera calibration method by a processor.

[0028] Compared with existing technologies, the present invention has the following advantages: Based on Zhang Zhengyou's camera calibration method, the present invention uses deep learning target detection technology to detect corner points in images, and designs a corner point coordinate correction method based on homography matrix according to the mathematical principle of camera calibration. The deep learning model, which has undergone extensive learning, can robustly extract the corner point positions in the image, overcoming the problems of low resolution, low signal-to-noise ratio and image blurring caused by fixed focal length in infrared images. It has the advantages of being robust and accurate, and provides a stable and reliable calibration method for commonly used fixed-focus infrared cameras. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of the network structure of the improved YOLOv8 target detection model in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of the network structure of the Local Window Attention Mechanism (LWA) module in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of an ideal imaging model of the camera in an embodiment of the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] like Figure 1 As shown, the thermal infrared fixed-focus camera calibration method based on deep neural networks in this embodiment includes the following steps: S1, using a pre-trained target detection model based on deep neural networks to identify the pixel coordinates of the corner points in the checkerboard image (e.g., ...). Figure 1 (as shown by the circles in the image); S2, use the homography matrix to perform a global correction on the pixel coordinates of all corner points to obtain the corrected pixel coordinates (e.g., ...). Figure 1 (As shown by the "+" in the middle); S3, based on the corrected pixel coordinates of each corner point, use Zhang Zhengyou's camera calibration method to calibrate the camera and obtain the camera's intrinsic parameter matrix and distortion coefficients. For example... Figure 1 As shown, in this embodiment, before step S1, there is also a step of training the target detection model using manually labeled checkerboard images.

[0035] As an optional implementation method, such as Figure 2 As shown, the object detection model in this embodiment is an improved version of the YOLOv8 object detection model. This improved model adds a cascaded Local Window Attention (LWA) module to the output of the backbone network of the YOLOv8 object detection model to enhance the network's ability to extract local image information. The YOLOv8 object detection model mainly consists of a backbone network, a neck network, and a detection network. Figure 2 Apart from the Local Window Attention (LWA) module, the other modules are existing modules of the YOLOv8 object detection model, so they will not be described in detail here.

[0036] like Figure 3 As shown, in this embodiment, the Local Window Attention Mechanism (LWA) module processes the input features from the output of the backbone network as follows: First, the input feature map is divided into non-overlapping windows. Then, each window is sequentially concatenated on the depth channel. Attention mechanism calculations are performed with the window center as the anchor point. Finally, the calculation result is returned to the position of the anchor point in the original feature map. This method combines the feature extraction capability of the attention mechanism with the reduction of model computation by using non-overlapping window division.

[0037] Ideal pixel coordinates, i.e., pixel coordinates after corner correction. World coordinates can be obtained through a homography matrix H. To obtain, that is:

[0038] ,

[0039] In the above formula, These are the pixel coordinates after corner correction. , , , , , , and Homography matrix Elements in the homography matrix The degree of freedom is 8. From the above formula, we can obtain the following two equations:

[0040] ,

[0041] To solve for the eight parameters of the homography matrix, only eight equations need to be established using the coordinates of four pairs of corner points. However, in practice, the calibration plate provides far more than four corner points, meaning the number of equations exceeds the number of unknowns. Since the camera parameters are fixed, the homography matrix H obtained by solving the eight-grid equations constructed from the coordinates of any four corner points and their world coordinates should be the same. Any different solutions would only be due to errors in the corner point coordinates. Therefore, to correct for these errors, step S2 in this embodiment uses the homography matrix to perform an overall correction on all corner point coordinates, resulting in the corrected corner point coordinates, including:

[0042] S2.1, based on the pixel coordinates of the n corner points in the identified chessboard image. and their corresponding world coordinates Substituting into the equation shown below, we obtain a total of 2n equations:

[0043] ,

[0044] In the above formula, , , , , , , and Homography matrix The elements in the set, and have:

[0045] ;

[0046] S2.2, Solve the 2n equations using the least squares method to obtain the homography matrix. Elements in;

[0047] S2.3, homography matrix Substituting the elements in the formula into the overall formula, we can obtain the corrected corner coordinates by correcting all corner coordinates:

[0048] ,

[0049] In the above formula, These are the pixel coordinates after corner correction.

[0050] Figure 4 This is a schematic diagram of an ideal imaging model for a camera. This model illustrates the geometric relationships between the world coordinate system, camera coordinate system, image coordinate system, and pixel coordinate system, with the symbols representing coordinate axes and point coordinates, respectively. World coordinate system The real point is... Corner point For the camera coordinate system. Based on... Figure 1 The ideal imaging model shown can be described as follows: The process of projecting 3D world coordinate points in the world coordinate system to 2D pixel coordinate points in the pixel coordinate system can be described as follows:

[0051] ,

[0052] In the above formula, As a scale factor, These are the coordinates of a 2D pixel point in a pixel coordinate system. and Here is the extrinsic parameter matrix of the camera. Let be a 3×3 camera intrinsic parameter matrix, and we have:

[0053] ,

[0054] in, and The focal length of the camera (in pixels). and This represents the principal point position of the camera. Furthermore, due to errors caused by manufacturing processes and other factors, image distortion will occur. The pixel position relationship before and after distortion is as follows:

[0055] ,

[0056] in, and These are the pixel coordinates before and after distortion, respectively. , , The radial distortion coefficient is... , This refers to the tangential distortion coefficient. Camera calibration essentially involves using a large number of image corner points... and its corresponding real point The mapping between them estimates the intrinsic parameter matrix of the camera. and distortion coefficient , , , , Currently, the most widely used method is Zhang Zhengyou's camera calibration method. This method models camera calibration as an optimization problem, with the optimization objective being to minimize the reprojection error. In this embodiment, step S3, using Zhang Zhengyou's camera calibration method to obtain the camera's intrinsic parameter matrix and distortion coefficients based on the corrected pixel coordinates of each corner point, refers to solving the optimization problem shown in the following equation to finally obtain the camera's intrinsic parameter matrix and distortion coefficients:

[0057] ,

[0058] In the above formula, The number of chessboard images. This represents the number of corner points in each chessboard image. Let J be the corrected pixel coordinates of the j-th corner point in the i-th chessboard image. Let J be the coordinates of the reprojection point of the j-th corner point in the i-th chessboard image. This is the intrinsic parameter matrix of the camera. Let be the rotation matrix of the i-th chessboard image. Let be the translation vector of the i-th chessboard image. Let j be the world coordinates of the j-th corner point in the i-th chessboard image. The distortion coefficients include radial and tangential distortion coefficients. The camera intrinsic parameters and distortion coefficients are continuously optimized during the iterative process of minimizing the above expression until the termination condition is met, at which point the solution is complete.

[0059] To verify the thermal infrared fixed-focus camera calibration method based on deep neural networks in this embodiment, experimental verification was conducted. The specific steps included: 1) Data acquisition. To obtain high-contrast infrared images, a portable heating platform was used to heat a calibration plate (with a checkerboard image) made of materials with different emissivity. The size of the calibration plate was smaller than the heating area, so it was covered with heat-insulating material to prevent exposure during infrared imaging. The imaging device used was a thermal infrared camera on the DJI Mavic 3T drone gimbal. 2) Data partitioning. The acquired data was divided into two parts: a training set and a test set, and labels were manually created. The training set was used to train a deep learning object detection model, which could then be used in a wide range of checkerboard corner detection tasks. The test set was used to verify the performance of the object detection model and the subsequent camera calibration process. In this embodiment, 200 training images and 50 test images were acquired, with the sampled spatial locations having the same distribution range. 3) Model training. The improved YOLOv8 object detection model was trained using the YOLOv8n.pt provided by Ultralytics. The training consisted of 200 epochs with a batch size of 16, using the SGD optimizer. Mosaic data augmentation was disabled in the last 10 epochs to allow the network to converge. Furthermore, the original YOLOv8 model was trained in the same manner for experimental comparison to demonstrate the effectiveness of the improvement. 4) Corner Detection: A pre-trained deep network detection model was used to detect corners in the test images, obtaining rectangular boxes containing the corners. The center of these boxes was used as the initial corner coordinates. 5) Corner Coordinate Correction: The proposed homography matrix correction method was used to correct the initial corner coordinates, resulting in corrected coordinates. 6) Camera Calibration: The corrected corner coordinates were used for camera calibration to obtain the final camera intrinsic parameters, distortion coefficients, and reprojection error.

[0060] In the corner detection results, our evaluation criteria mainly focus on the corner false negative rate, the number of false alarms, and the number of rejected images. The false negative rate and the number of false alarms reflect the robustness of the detection method. The number of rejected images reflects the impact of the detection method on the camera calibration task. In the camera calibration results, we used the root mean square error (RMSE) and the mean reprojection error (MRE), calculated as follows:

[0061] ,

[0062] ,

[0063] In the above formula, The number of chessboard images. This represents the number of corner points in each chessboard image. Let J be the corrected pixel coordinates of the j-th corner point in the i-th chessboard image. Let be the coordinates of the reprojection point of the j-th corner in the i-th checkerboard image. The maximum reprojection error and the standard deviation of the reprojection error were also calculated. These metrics comprehensively reflect the accuracy of the camera calibration. The corner detection results of our method and camera calibration tools in OpenCV and MATLAB on 50 test images are shown in Table 1.

[0064] Table 1 Corner detection results of different methods

[0065]

[0066] In Table 1, the improved YOLOv8 is the improved model of the YOLOv8 object detection model in this embodiment. As shown in Table 1, the original YOLOv8 model detected 7 false corners in one image, making it unusable for camera calibration. In contrast, the improved YOLOv8 successfully detected all corners without any false alarms. This demonstrates that the Local Window Attention (LWA) module effectively improves the robustness of corner detection.

[0067] After comparing the corner detection results, camera calibration was performed using corner coordinates obtained by different methods. To verify the effectiveness of our proposed homography matrix correction method, camera calibration was performed using both the initial and corrected corner coordinates, corresponding to methods 1 and 2, respectively. To verify the effectiveness of the network model improvement, the original YOLOv8 model trained under the same conditions was used for the same tests. The final calibration error evaluation metrics for the six methods are shown in Table 2, and the camera parameters and distortion coefficients obtained by the six methods are shown in Table 3.

[0068] Table 2. Calibration errors of different methods

[0069]

[0070] Table 3. Calibration results of different methods

[0071]

[0072] See Figure 2 and Figure 3 As can be seen, our proposed method achieves optimal values ​​across all four evaluation metrics. Under otherwise identical conditions, the method using the original YOLOv8 model is suboptimal across all four metrics. This demonstrates that the Local Window Attention (LWA) module effectively improves the accuracy of corner detection. Specifically, our method (number 2) obtains the principal point position... The parameters are closest to the center position of the image (320, 256) (image resolution is 640×512). It can be seen that the thermal infrared fixed-focus camera calibration method based on deep neural networks in this embodiment can effectively improve the robustness of corner detection. The corner coordinate correction method based on the homography matrix can effectively improve the corner detection accuracy and reduce the reprojection error in camera calibration. Actual experimental results based on thermal infrared camera calibration show that the thermal infrared fixed-focus camera calibration method based on deep neural networks in this embodiment has significantly improved the accuracy of camera calibration compared to existing mainstream camera calibration tools (OpenCV and MATLAB), providing a practical, stable, and reliable calibration method for commonly used fixed-focus thermal infrared cameras.

[0073] Furthermore, this embodiment also provides a thermal infrared fixed-focus camera calibration system based on deep neural networks, including:

[0074] The corner recognition program unit is used to identify the coordinates of corner points in a chessboard image using a pre-trained object detection model based on a deep neural network.

[0075] The corner coordinate correction program unit is used to perform a global correction on the coordinates of all corner points using the homography matrix to obtain the corrected corner coordinates.

[0076] The parameter calibration program unit is used to calibrate the camera using the Zhang Zhengyou camera calibration method based on the corrected corner coordinates to obtain the camera's intrinsic parameter matrix and distortion coefficients.

[0077] Furthermore, this embodiment also provides a thermal infrared fixed-focus camera calibration system based on deep neural networks, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the thermal infrared fixed-focus camera calibration method based on deep neural networks.

[0078] In addition, this embodiment also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the deep neural network-based thermal infrared fixed-focus camera calibration method by a processor.

[0079] In addition, this embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the deep neural network-based thermal infrared fixed-focus camera calibration method by a processor.

[0080] Those skilled in the art will understand that the technical solutions provided by the embodiments of this application may be in the form of a method, system, or computer program product. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create an implementation for the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0081] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A calibration method for a thermal infrared fixed-focus camera based on a deep neural network, characterized in that, The process includes the following steps: S1, using a pre-trained object detection model based on a deep neural network to identify the pixel coordinates of the corner points in the checkerboard image; S2, using a homography matrix to correct the pixel coordinates of all corner points to obtain the corrected pixel coordinates; S3, using the Zhang Zhengyou camera calibration method to calibrate the camera based on the corrected pixel coordinates of each corner point to obtain the camera's intrinsic parameter matrix and distortion coefficients. Step S2 uses the homography matrix to perform a global correction on the coordinates of all corner points, resulting in the corrected corner point coordinates, including: S2.1, based on the pixel coordinates of the n corner points in the identified chessboard image. and their corresponding world coordinates Substituting into the equation shown below, we obtain a total of 2n equations: ， In the above formula, , , , , , , and Homography matrix The elements in the set, and have: ； S2.2, Solve the 2n equations using the least squares method to obtain the homography matrix. Elements in; S2.3, homography matrix Substituting the elements in the formula into the overall formula, we can obtain the corrected corner coordinates by correcting all corner coordinates: ， In the above formula, These are the pixel coordinates after corner correction; In step S3, the camera is calibrated using Zhang Zhengyou's camera calibration method based on the corrected pixel coordinates of each corner point to obtain the camera's intrinsic parameter matrix and distortion coefficients. This involves solving the optimization problem shown in the following equation to ultimately obtain the camera's intrinsic parameter matrix and distortion coefficients: ， In the above formula, The number of chessboard images. This represents the number of corner points in each chessboard image. Let J be the corrected pixel coordinates of the j-th corner point in the i-th chessboard image. Let J be the coordinates of the reprojection point of the j-th corner point in the i-th chessboard image. This is the intrinsic parameter matrix of the camera. Let be the rotation matrix of the i-th chessboard image. Let be the translation vector of the i-th chessboard image. Let j be the world coordinates of the j-th corner point in the i-th chessboard image. These are the distortion coefficients, including radial distortion coefficients and tangential distortion coefficients.

2. The thermal infrared fixed-focus camera calibration method based on deep neural networks according to claim 1, characterized in that, The object detection model is an improved version of the YOLOv8 object detection model. The improved YOLOv8 object detection model adds a cascaded Local Window Attention (LWA) module to the output of the backbone network of the YOLOv8 object detection model to improve the network's ability to extract local information from the image.

3. The thermal infrared fixed-focus camera calibration method based on deep neural networks according to claim 2, characterized in that, The Local Window Attention Mechanism (LWA) module processes the input features from the output of the backbone network as follows: first, the input feature map is divided into non-overlapping windows; then, each window is sequentially stitched together on the depth channel; attention mechanism calculation is performed with the center of the window as the anchor point; and finally, the calculation result is returned to the position of the anchor point in the original feature map.

4. The thermal infrared fixed-focus camera calibration method based on deep neural networks according to claim 3, characterized in that, Before step S1, there is also a step of training the object detection model using manually labeled checkerboard images.

5. A calibration system for a thermal infrared fixed-focus camera based on a deep neural network, characterized in that, include: The corner recognition program unit is used to identify the coordinates of corner points in a chessboard image using a pre-trained object detection model based on a deep neural network. The corner coordinate correction program unit is used to perform a global correction on the coordinates of all corner points using the homography matrix to obtain the corrected corner coordinates. The parameter calibration program unit is used to calibrate the camera using the Zhang Zhengyou camera calibration method based on the corrected corner coordinates to obtain the camera's intrinsic parameter matrix and distortion coefficients; The thermal infrared fixed-focus camera calibration system is used to be programmed or configured to perform the deep neural network-based thermal infrared fixed-focus camera calibration method according to any one of claims 1 to 4.

6. A calibration system for a thermal infrared fixed-focus camera based on a deep neural network, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the deep neural network-based thermal infrared fixed-focus camera calibration method according to any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the deep neural network-based thermal infrared fixed-focus camera calibration method described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the deep neural network-based thermal infrared fixed-focus camera calibration method described in any one of claims 1 to 4.

Images