Ship thermal infrared wake detection method based on rhombus prediction frame YOLO model

Abstract

The invention discloses a ship thermal infrared wake detection method based on a rhombus prediction frame YOLO model, and the method comprises the steps: replacing a conventional rectangular frame with a rotatable rhombus prediction frame, and positioning a wake candidate region through the center coordinate of a rhombus, the length of a diagonal line, and a rotation angle; thirdly, Gaussian filtering denoising is carried out on the candidate area, potential line segments are detected in combination with line support domain growth and rectangular fitting, effective line segments are screened through the false alarm number and the Euclidean distance, and the image is segmented based on gray average and variance binaryzation; and finally, extracting a continuous wake contour through edge detection, and fitting and optimizing morphological parameters in combination with a recursive broken line. By utilizing the method, the wake target on the thermal infrared remote sensing image can be quickly detected, the wake contour can be accurately fitted, the detailed form information can be acquired, and the method is particularly suitable for the application requirement of real-time monitoring of the sea surface moving target.

Description

Technical Field

[0001] The present invention belongs to the field of infrared photoelectric remote sensing detection technology, and in particular relates to a ship thermal infrared wake detection method based on a diamond prediction box YOLO model. Background Art

[0002] Current space-based remote sensing technologies for maritime imaging primarily include synthetic aperture radar (SAR), visible light, and infrared imaging. Thermal infrared imaging is more adaptable to the environment than visible light, making nighttime imaging unaffected. Its passive radiation reception principle makes it more concealed and safer than synthetic aperture radar. When ships navigate in uniform seawater, the surrounding seawater temperature rises due to heat release from the propulsion system, propeller water injection, and friction. Buoyancy forces the warmer water to rise to the surface, forming a thermal wake. This thermal wake is unique to ships, has a large range, lasts for a long time, and is easily detected by infrared detectors. Therefore, thermal infrared wakes are a crucial feature for ship target detection.

[0003] Since there are relatively few studies on thermal infrared wake detection and it started relatively late, most detection methods are derived from wake detection in synthetic aperture radar images, mainly based on neural network models and image analysis algorithms.

[0004] For example, the Chinese patent document with publication number CN119723476A discloses an infrared ship wake detection method and system that combines deformable convolution with an attention mechanism. Based on the improved wake detection model of yolov8, it comprehensively considers the image noise problem and small target detection problem under complex sea conditions, and proposes a deformable convolution, attention mechanism and multi-scale feature fusion module to improve the model detection performance, thereby improving the accuracy of infrared ship wake detection.

[0005] Chinese patent document CN107644414A discloses a sea surface wake detection method based on constant statistics and Radon transform, comprising: acquiring an infrared wake image using an infrared detector; performing non-uniform correction on the infrared wake image to obtain an image; performing maximum median filtering and template filtering on the corrected image in sequence to obtain an image; normalizing the filtered image and then performing threshold segmentation to binarize the image and highlight the target wake; and transforming the wake image into another parameter domain in the form of a line integral using Radon transform.

[0006] Traditional neural network detection models typically produce a rectangular region encompassing the target. This region, in addition to the wake, often covers a significant portion of the ocean surface, making it difficult to accurately match the actual wake contour. Image analysis algorithms are also sensitive to noise when processing complex scenes, making them susceptible to interference and resulting in a high false detection rate. Summary of the Invention

[0007] The present invention provides a ship thermal infrared wake detection method based on the diamond prediction box YOLO model, which can quickly detect wake targets on thermal infrared remote sensing images, accurately fit the wake contour, and obtain detailed morphological information (width, length, heading angle, curvature, inflection point). It is particularly suitable for the application requirements of real-time monitoring of moving targets on the sea surface.

[0008] A ship thermal infrared wake detection method based on the diamond prediction box YOLO model includes the following steps:

[0009] (1) Improve the YOLO model by using a rotatable diamond prediction box instead of a rectangular box;

[0010] (2) After pre-training the improved YOLO model, the thermal infrared remote sensing image is input into the improved YOLO model to preliminarily locate the diamond candidate area of ​​the trail;

[0011] (3) Using the adaptive non-maximum suppression method based on target density to filter the repeatedly detected tail areas, the selected diamond candidate areas are obtained;

[0012] (4) Using the selected diamond candidate area as a mask, the original image is segmented to extract the diamond-shaped sea surface area containing the thermal infrared trail;

[0013] (5) Apply Gaussian filtering to denoise the segmented diamond-shaped sea surface area, and detect all potential wake segments through the line support region growing algorithm and minimum rectangle fitting within the area;

[0014] (6) For each potential trail segment, double screening is performed by calculating the false alarm number NFA value and the Euclidean distance, and the valid trail segments are obtained after screening;

[0015] (7) Use the tail grayscale mean and variance to perform binary segmentation on all valid tail segments to obtain the tail and background;

[0016] (8) Fine edge detection to obtain the pixel-level boundary of the trail;

[0017] (9) Recursive broken line fitting is used to reconstruct the complete continuous curve shape of the wake and obtain the final thermal infrared wake detection result.

[0018] In step (2), when pre-training the improved YOLO model, the images in the training set are annotated using rotatable diamonds.

[0019] In step (2), the diamond candidate area of ​​the trail is located by the center point coordinates, diagonal length and rotation angle of the diamond.

[0020] In step (2), when pre-training the improved YOLO model, the total loss function is as follows:

[0021] L=λ cls L cls +λ box L box +λ dfl L dfl ;

[0022] Where, the default hyperparameter λ is cls =λ box =λ dfl =1;L cls For classification and target loss, VarifocalLoss is used as the loss function; L dfl is the distribution focus loss; L box The positioning loss for the rotatable diamond prediction box is as follows:

[0023]

[0024] ρ 2 ((x,y),(x * ,y * ))=(x p -x t ) 2 +(y p -y t ) 2 ;

[0025]

[0026] Where, IoU rot (b,b * ) is the area intersection and union ratio of the rotated diamond prediction box and the true box; s * is the target score, which reflects the importance of the positive sample and is 0 for the negative sample; ρ((x,y),(x * ,y * )) is the Euclidean distance between the center of the diamond prediction box and the center of the real box; (x p ,y p ) is the coordinate of the center point of the diamond prediction box, (x t ,y t ) is the coordinate of the center point of the real diamond box; c is used to normalize the center point distance loss, which is defined as the distance between the two farthest points on the diamond prediction box and the real box; a p and b p are the lengths of the diagonal lines of the diamond prediction boxes, a t and b t are the lengths of the diagonals of the true diamond boxes, respectively. v is used to measure the difference in the ratio of the diagonals. The larger v is, the greater the penalty is. angle is a hyperparameter;

[0027] (1-cos(θp -θ t )) is used to measure the loss of the rotation angle between the rotatable prediction box and the real box, θ p and θ t are the rotation angles between the main and horizontal directions of the diamond box prediction box and the real diamond box, respectively.

[0028] The specific process of step (3) is:

[0029] First, the distribution of diamond candidate regions is analyzed and the density of the target is estimated. According to the density estimation result, the overlap threshold of non-maximum suppression is dynamically set. Then, the diamond candidate regions exceeding the overlap threshold are filtered to obtain the filtered diamond candidate regions.

[0030] In step (5), Gaussian filtering uses a weight matrix that follows a two-dimensional Gaussian distribution to perform weighted averaging on each pixel in the segmented diamond-shaped sea area image:

[0031]

[0032] Where x and y are the distances between the pixel and the kernel center, and σ is the standard deviation of the Gaussian distribution.

[0033] In step (5), all potential trail segments are detected by using the line support region growing algorithm and the minimum rectangle fitting within the region. The specific process is as follows:

[0034] First, starting from a pixel point P(x,y) on the diamond-shaped sea area image, calculate its horizontal and vertical gradient g x (x,y) and g y (x, y), then calculate the gradient direction θ(x, y) and gradient magnitude G(x, y);

[0035] The gradient direction of the pixel is used as the gradient direction of the area to be grown, and then the gradient directions of the adjacent pixels are calculated. If the difference between the gradient direction of the adjacent pixel and the gradient direction of the area to be grown is lower than the set threshold, the adjacent pixel is included in the area, and the gradient direction of the area is iteratively updated. The iteration is continued until all the adjacent pixels of the area to be grown do not meet the requirements. All the areas to be grown where the iteration is stopped are transformed into candidate areas.

[0036] A minimum bounding rectangle is used to fit each candidate region. This minimum bounding rectangle is the potential tail segment of the preliminary fitting. The major axis of the rectangle corresponds to the direction of the fitted straight line segment, the minor axis corresponds to the direction perpendicular to the gradient direction, and the center of the rectangle represents the center of mass of the region.

[0037] In step (7), when all valid trail segments are binarized using the trail grayscale mean and variance, the threshold range formula for segmentation is as follows:

[0038] T min =T average -kσ;

[0039] T max =T average +kσ;

[0040] Where, T min and T max is the lower and upper thresholds of the segmentation range, T average is the grayscale mean, σ is the standard deviation, and k is the coefficient for controlling the size of the range.

[0041] In step (9), the recursive broken line fitting is used to reconstruct the complete continuous curve shape of the tail. The specific process is as follows:

[0042] First, select the first and last points of the curve detected by edge detection and connect them into a line segment;

[0043] Then calculate the vertical distance from other points on the curve to this line segment and find the point with the largest distance. If this maximum distance exceeds the set precision parameter ε, retain the point and use it as the new dividing point to divide the curve into two segments, forming a broken line.

[0044] Repeat the above operation until the distance from all points on the curve to the broken line is less than ε; thus obtaining the final thermal infrared wake detection result.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] In view of the morphological characteristics of trails, the present invention innovatively adopts a rotatable diamond prediction box to replace the rectangular prediction box in the traditional YOLO series models, and at the same time adds a pixel-level image analysis algorithm for post-processing. It can adapt to the geometric characteristics of trails more finely when fitting slender, broken lines or sharp bends, and significantly reduces the positioning error caused by the rotating rectangular box when processing corners and changing curvatures. It has the advantages of high efficiency, accuracy and intuitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a method for detecting thermal infrared wakes of ships based on the diamond prediction box YOLO model according to an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the output results of key steps in an embodiment of the present invention.

[0049] Figure 3 Schematic diagram of positioning parameters of a rotatable diamond prediction box in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It should be noted that the following examples are intended to facilitate understanding of the present invention and do not have any limiting effect on the present invention.

[0051] Aiming at the problem that the current thermal infrared wake detection algorithm or detection results contain too much sea surface area or are greatly affected by noise, and cannot accurately fit the wake and obtain detailed information, which makes subsequent research difficult to advance, this paper proposes a ship thermal infrared wake detection method based on the diamond prediction box YOLO model. Figure 1 The specific implementation steps are as follows:

[0052] (1) The rectangular prediction box of the traditional YOLO series model is replaced by a rotatable diamond prediction box. At the same time, the training set annotation method also needs to be improved synchronously. The trained YOLO model is used to detect the sea surface thermal infrared remote sensing image to preliminarily locate the wake area. The improvement of the rotatable diamond prediction box is applicable to all models in the YOLO series that use rectangular prediction boxes for prediction. Taking the latest YOLOv12 as an example, the total loss function of the new model is:

[0053] L=λ cls L cls +λ box L box +λ dfl L dfl ;

[0054] Default hyperparameter λ cls =λ box =λ dfl =1, where L cls For classification and target loss, VarifocalLoss is used as the loss function, L dfl For the distribution focus loss, both items are inherited from the YOLOv12 model. The specific formula is:

[0055]

[0056] in, is the logit output of the i-th prediction box, q i ∈[0,1] corresponds to the quality score of the real box, l i ∈{0,1} is the positive and negative sample label, p is the discrete probability distribution predicted by the model, t is the true continuous label value, α and γ are hyperparameters, and CE(p,c) is the cross entropy loss for category c.

[0057] This paper focuses on improving the L in the model loss function box The loss function changes the original rectangular prediction box positioning loss to a rotatable diamond prediction box positioning loss. The specific loss function is as follows:

[0058]

[0059] ρ 2 ((x,y),(x * ,y * ))=(x p -x t ) 2 +(y p -y t ) 2 ;

[0060]

[0061] Among them, IoU rot (b,b * ) is the area intersection and union ratio of the rotated diamond prediction box and the true box; s * is the target score, which reflects the importance of the positive sample and is 0 for the negative sample; ρ((x,y),(x * ,y * )) is the Euclidean distance between the center of the prediction box and the center of the real box; c is used to normalize the center point distance loss, which is defined as the distance between the two farthest points on the diamond prediction box and the real box; a and b are the lengths of the two diagonals of the diamond prediction box, respectively, and v is used to measure the difference in the ratio of the diagonals. The larger v is, the greater the penalty is; λ angle is a hyperparameter; (1-cos(θ p -θ t )) is used to measure the loss of the rotation angle between the rotatable prediction box and the real box, and θ is the angle between the main diagonal of the diamond box and the horizontal direction.

[0062] The above is an introduction to the loss function of the improved model. The trained improved model is used to detect thermal infrared images and obtain diamond-shaped areas containing tails.

[0063] (2) Use the adaptive non-maximum suppression method based on target density to filter the tail area of ​​repeated detections. First, analyze the distribution of the detection frame and estimate the density of the target. According to the density estimation result, the overlap threshold of non-maximum suppression is dynamically set. The overlap threshold refers to a key parameter in the non-maximum suppression process to determine whether the detection frames overlap too much and therefore need to suppress other detection frames. When the target density is high, increase the overlap threshold so that only the detection frames with a high degree of overlap are suppressed; when the target density is low, lower the overlap threshold to relax the restriction on the overlap of the detection frames and avoid mistakenly suppressing adjacent real targets.

[0064] (3) Using the filtered prediction box as a mask, the original image is segmented to extract the diamond-shaped sea surface area containing the thermal infrared trail.

[0065] (4) Gaussian blur is used to denoise the diamond-shaped sea surface area containing thermal infrared trails. After analyzing the inherent noise characteristics of the image, it is found that the main factor that has a greater impact on the algorithm is the grid stripe noise generated by the infrared payload push sweep. Because of its periodicity and repetitiveness, Gaussian blur is used to denoise the image. Gaussian blur uses a weight matrix (Gaussian kernel) that follows a two-dimensional Gaussian distribution to perform weighted averaging on each pixel of the image:

[0066]

[0067] Where x and y are the distances between the pixel and the kernel center, and σ is the standard deviation of the Gaussian distribution.

[0068] (5) All potential tail line segments are detected in the preprocessed image by using the line support region growing algorithm and the minimum rectangle fitting in the region. First, starting from a pixel point P(x,y) on the image, its horizontal and vertical gradients g are calculated. x (x,y) and g y (x,y), the calculation formulas are:

[0069]

[0070] Where i(x,y) represents the grayscale value at the pixel point P(x,y). The gradient direction θ(x,y) and gradient magnitude G(x,y) are then calculated using the following formula:

[0071]

[0072] The gradient direction of the pixel is used as the gradient direction of the region to be grown, and then the gradient direction of the adjacent pixels is calculated. If the difference between the gradient direction of the adjacent pixel and the gradient direction of the region is lower than the set threshold, the adjacent pixel is included in the region, and the gradient direction of the region is iteratively updated. The gradient direction of the region can be expressed according to the following formula:

[0073]

[0074] The iteration continues until all adjacent pixels in the region to be grown do not meet the requirements. For each candidate segment, a minimum bounding rectangle is used to fit the region. The major axis of the rectangle represents the main direction of the segment, that is, the direction of the fitted straight line segment; the minor axis represents the direction perpendicular to the gradient direction; the center of the rectangle represents the center of mass of the region, and the corner points of the rectangle define the boundary of the rectangle.

[0075] (6) By calculating the number of false alarms (NFA) and the Euclidean distance of the detected line segments, the noise is filtered out and the true trails are selected. NFA is based on hypothesis testing and measures the probability of observing a line segment in a purely random background. The formula is as follows:

[0076] NFA=N tests ·P;

[0077] Among them, N tests is the number of all possible candidate line segments in the image, and P is the probability of randomly generating a line segment that matches the parameters (length, direction, gradient consistency, etc.) of the current line segment in a pure noise image. Detection results with NFA ≥ 1 are considered noise and discarded. The remaining line segments are then sorted according to Euclidean distance, and a threshold is set. Only line segments with a length greater than the threshold are considered true trails.

[0078] (7) Binarize the image based on the grayscale mean and variance of the detected trails. Linearly interpolate the detected trails from the starting point to the end point, extract the grayscale values ​​of the pixels, and calculate the mean and standard deviation as the threshold for image segmentation. The threshold range is calculated using the following formula:

[0079] T min =T average -kσ;

[0080] T max =T average +kσ;

[0081] Among them, k is used to control the range size, T min and T max are the lower and upper thresholds of the segmentation range, T average is the grayscale mean, and σ is the standard deviation.

[0082] (8) For the image that has been binarized and segmented, edge detection is used to extract the complete tail and fit the curve. The image after segmentation has a relatively obvious boundary, so the edge detection method can be used to extract the complete tail. The main steps are Gaussian preprocessing, gradient calculation, non-maximum suppression, double threshold detection and edge connection.

[0083] (9) Perform curve fitting on the results after edge detection, and then remove noise based on the length of the curve. The specific operation is: first select the first and last points of the curve detected by the edge, connect them into a line segment, and then calculate the vertical distance from other points on the curve to this line segment, and find the point with the largest distance. If this maximum distance exceeds the set accuracy parameter ε, retain the point and use it as the new dividing point to divide the curve into two segments, forming a broken line. Repeat the above operation until the distance from all points on the curve to the broken line is less than ε. The detection result at this time is the final thermal infrared trail.

[0084] like Figure 2 The following image shows the output of key steps in the algorithm's complete run on sample images, including single-tail images, multiple-tail images, straight-line tail images, and curved-tail images. The output of the algorithm in the first stage shows the diamond prediction box detection result, the image segmentation result, and the final tail result after the second stage of detection.

[0085] The method of the present invention improves accuracy by combining a rotating diamond frame with pixel-level edge detection. It can adapt to the geometric characteristics of the wake more finely when fitting slender, broken lines or sharp bends, and significantly reduces the positioning error caused by the rotating rectangular frame when processing corners and changing curvature. It can support real-time monitoring of the sea surface and efficient target tracking of ship situation awareness systems.

[0086] like Figure 3 The figure shows the positioning parameters of the rotatable diamond prediction box, where (x, y) is the coordinate of the center point of the prediction box, a and b are the lengths of the diagonal of the prediction box, and θ is the rotation angle between the diamond prediction box and the horizontal direction.

[0087] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A ship thermal infrared wake detection method based on the diamond prediction box YOLO model, characterized in that: The following steps are involved: (1) Improve the YOLO model by using a rotatable diamond prediction box instead of a rectangular box; (2) After pre-training the improved YOLO model, the thermal infrared remote sensing image is input into the improved YOLO model to preliminarily locate the diamond candidate area of ​​the trail; (3) Using the adaptive non-maximum suppression method based on target density to filter the repeatedly detected tail areas, the selected diamond candidate areas are obtained; (4) Using the selected diamond candidate area as a mask, the original image is segmented to extract the diamond-shaped sea surface area containing the thermal infrared trail; (5) Apply Gaussian filtering to denoise the segmented diamond-shaped sea surface area, and detect all potential wake segments through the line support region growing algorithm and minimum rectangle fitting within the area; (6) For each potential trail segment, double screening is performed by calculating the false alarm number NFA value and the Euclidean distance, and the valid trail segments are obtained after screening; (7) Use the tail grayscale mean and variance to perform binary segmentation on all valid tail segments to obtain the tail and background; (8) Fine edge detection to obtain the pixel-level boundary of the trail; (9) Recursive broken line fitting is used to reconstruct the complete continuous curve shape of the wake and obtain the final thermal infrared wake detection result.

2. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: In step (2), when pre-training the improved YOLO model, the images in the training set are annotated using rotatable diamonds.

3. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: In step (2), the diamond candidate area of ​​the trail is located by the center point coordinates, diagonal length and rotation angle of the diamond.

4. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: In step (2), when pre-training the improved YOLO model, the total loss function is as follows: L=λ cls L cls +λ box L box +λ dfl L dfl ; Where, the default hyperparameter λ is cls =λ box =λ dfl =1;L cls For classification and target loss, Varifocal Loss is used as the loss function; L dfl is the distribution focus loss; L box The positioning loss for the rotatable diamond prediction box is as follows: Where, IoU rot (b,b * ) is the area intersection and union ratio of the rotated diamond prediction box and the true box; s * is the target score, which reflects the importance of the positive sample and is 0 for the negative sample; ρ((x,y),(x * ,y * )) is the Euclidean distance between the center of the diamond prediction box and the center of the real box, (x p ,y p ) is the coordinate of the center point of the diamond prediction box, (x t ,y t ) is the coordinate of the center point of the real diamond box; c is used to normalize the center point distance loss, which is defined as the distance between the two farthest points on the diamond prediction box and the real box; a p and b p are the lengths of the diagonal lines of the diamond prediction boxes, a t and b t are the lengths of the diagonals of the true diamond boxes, respectively. v is used to measure the difference in the ratio of the diagonals. The larger v is, the greater the penalty is. angle is a hyperparameter; (1-cos(θ p -θ t )) is used to measure the loss of the rotation angle between the rotatable prediction box and the real box, θ p and θ t are the rotation angles between the diamond box prediction box and the real diamond box and the horizontal direction, respectively.

5. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: The specific process of step (3) is: First, the distribution of diamond candidate regions is analyzed and the density of the target is estimated. According to the density estimation result, the overlap threshold of non-maximum suppression is dynamically set. Then, the diamond candidate regions exceeding the overlap threshold are filtered to obtain the filtered diamond candidate regions.

6. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: In step (5), Gaussian filtering uses a weight matrix that follows a two-dimensional Gaussian distribution to perform weighted averaging on each pixel in the segmented diamond-shaped sea area image.

7. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: In step (5), all potential trail segments are detected by using the line support region growing algorithm and the minimum rectangle fitting within the region. The specific process is as follows: First, starting from a pixel point on the diamond-shaped sea surface area image, its horizontal and vertical gradients are calculated, and then the gradient direction and gradient amplitude are calculated; The gradient direction of the pixel is used as the gradient direction of the region to be grown, and then the gradient directions of the adjacent pixels are calculated. If the difference between the gradient direction of the adjacent pixel and the gradient direction of the region to be grown is lower than the set threshold, the adjacent pixel is included in the region, and the gradient direction of the region is iteratively updated. The iteration is continued until all the adjacent pixels of the region to be grown do not meet the requirements. All the regions to be grown where the iteration is stopped are transformed into candidate regions. A minimum bounding rectangle is used to fit each candidate region. This minimum bounding rectangle is the potential tail segment of the preliminary fitting. The major axis of the rectangle corresponds to the direction of the fitted straight line segment, the minor axis corresponds to the direction perpendicular to the gradient direction, and the center of the rectangle represents the center of mass of the region.

8. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: In step (7), when all valid trail segments are binarized using the trail grayscale mean and variance, the threshold range formula for segmentation is as follows: T min =T average -kσ; T max =T average +kσ; Where, T min and T max are the lower and upper thresholds of the segmentation range, T average is the grayscale mean, σ is the standard deviation, and k is the coefficient for controlling the size of the range.

9. The ship thermal infrared wake detection method based on the diamond prediction box YOLO model according to claim 1 is characterized in that: In step (9), the recursive broken line fitting is used to reconstruct the complete continuous curve shape of the tail. The specific process is as follows: First, select the first and last points of the curve detected by edge detection and connect them into a line segment; Then calculate the vertical distance from other points on the curve to this line segment and find the point with the largest distance. If this maximum distance exceeds the set precision parameter ε, retain the point and use it as the new dividing point to divide the curve into two segments, forming a broken line. Repeat the above operation until the distance from all points on the curve to the broken line is less than ε; thus obtaining the final thermal infrared wake detection result.

Images