A multi-directional gradient-based infrared dim small target detection method

Abstract

The application discloses a kind of infrared weak and small target detection methods based on multi-direction gradient, first, the gradient of each direction of original image is calculated using Facet model, then in each direction, along the direction gradient, find the crest and trough, according to the distance between current point and crest and trough, determine whether current point is target point, fuse the target point information of each direction to obtain the final target position candidate point;According to target position candidate point information, the target is enhanced, while suppressing background, calculate the saliency map of each direction, finally, the saliency map of each direction is fused to obtain the final saliency map, and the target is segmented according to the saliency map.The present application can be widely used in weak and small target detection of infrared image.

Description

Technical Field

[0001] This invention relates to a method for detecting weak infrared targets using multi-directional gradients, and relates to the field of digital image processing. Background Technology

[0002] Infrared weak target detection is an important topic in digital image processing, widely used in various fields, including infrared search systems, tracking systems, and aerospace defense systems. Due to the long imaging distance, weak targets, low signal-to-clutter ratio, and lack of shape and texture features, detection is challenging, and practical applications also place high demands on algorithm runtime. To achieve rapid detection of weak targets, this invention proposes an infrared weak target detection method based on multi-directional gradients. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an infrared weak target detection method based on multi-directional gradient.

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

[0005] A method for detecting weak infrared targets based on multi-directional gradients, the method comprising:

[0006] The Facet model is used to calculate the first-order partial derivatives in the neighborhood of each direction of the image to obtain the directional gradient; the target position in each direction is found based on the directional gradient, and the target position information in each direction is fused to determine the target candidate point; the target in each direction is enhanced while the background is suppressed to obtain the saliency map of the target candidate point in each direction, and the saliency maps of the target candidate point in each direction are fused to obtain the final saliency map; the target is extracted by threshold segmentation based on the final saliency map.

[0007] Furthermore, the Facet model is used to calculate the first-order partial derivatives in the 5x5 neighborhood of the image in each direction. The Facet model fits the 5x5 neighborhood image with a polynomial function, as shown in the following expression:

[0008] f(r, c) = b0 + b1r + b2c + b3(r) 2 -2)+b4rc+b5(c 2 -2)+b6(r 3 -17r / 5)+b7(r 2 -2)c+b8r(c 2 -2)+b9(c 3 -17c / 5)

[0009] Where r and c are the row and column coordinates within the 5x5 neighborhood image, respectively, and b i(i = 0, 1, ..., 9) are the fitting coefficients, and the first derivative at the horizontal direction angle α is:

[0010]

[0011] in,

[0012]

[0013] Furthermore, the b i We obtain it from the following formula:

[0014]

[0015] in, This represents the convolution operation, where I is the original image and w is the input image. i The following is an expression:

[0016]

[0017]

[0018]

[0019]

[0020] The above expressions can be used to calculate the derivative graphs f′0 and f′ of the image at four directions: 0°, 45°, 90°, and 135°. 45 f′ 90 f′ 135 .

[0021] Furthermore, the target location is located based on the directional gradient, and target location information from various directions is fused to determine candidate target points, specifically including:

[0022] In f′0, along the 0-degree direction of the current point, find the nearest peak to the current point to the left, and the value from the peak to the current point is monotonically decreasing, to obtain the distance dis_left. Find the nearest trough to the current point to the right, and the value from the trough to the current point is monotonically increasing, to obtain the distance dis_right. If at this time dis_left>0, dis_right>0, and the absolute value of the difference between dis_left and dis_right is within 3, then the current point is considered to be the target point in the 0-degree direction.

[0023] At f′ 45In the middle, along the 45-degree direction of the current point, find the nearest peak to the current point to the upper left, and the value of the peak to the current point is monotonically decreasing, to obtain the x-direction distance dis_left and the y-direction distance dis_up. Find the nearest trough to the current point to the lower right, and the value of the trough to the current point is monotonically increasing, to obtain the x-direction distance dis_right and the y-direction distance dis_down. If at this time (dis_left+dis_up)>0 and (dis_right+dis_down)>0, and the absolute value of the difference between (dis_left+dis_up) and (dis_right+dis_down) is within 3, then the current point is considered to be the target point in the 45-degree direction.

[0024] At f′ 90 In the process, along the 90-degree direction from the current point, find the nearest peak upwards where the value from the peak to the current point is monotonically decreasing, and obtain the y-direction distance dis_up. Find the nearest trough downwards where the value from the trough to the current point is monotonically increasing, and obtain the y-direction distance dis_down. If dis_up > 0 and dis_down > 0 at this time, and the absolute value of the difference between dis_up and dis_down is within 3, then the current point is considered to be the target point in the 90-degree direction.

[0025] At f′ 135 In the middle, along the 135-degree direction of the current point, find the nearest peak to the current point in the upper right direction, and the value of the peak to the current point is monotonically decreasing, to obtain the x-direction distance dis_right and the y-direction distance dis_up. Find the nearest trough to the current point in the lower left direction, and the value of the trough to the current point is monotonically increasing, to obtain the x-direction distance dis_left and the y-direction distance dis_down. If at this time (dis_right+dis_up)>0 and (dis_left+dis_down)>0, and the absolute value of the difference between (dis_right+dis_up) and (dis_left+dis_down) is within 3, then the current point is considered to be the target point in the 135-degree direction.

[0026] If the current point is a target point in the directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees, then the point is considered a candidate target point.

[0027] Furthermore, the target in each direction is enhanced while the background is suppressed, resulting in target point saliency maps for each direction. These saliency maps are then fused to obtain the final saliency map. Specifically, this includes:

[0028] At four directions (0°, 45°, 90°, 135°), for each candidate target point, based on the peak and trough positions obtained in the previous step, calculate the standard deviation s1 within the 3x3 neighborhood of the peak position, the standard deviation s2 within the 3x3 neighborhood of the trough position, and the difference d between the gradient values ​​at the peak and trough positions. This yields the saliency map t of the candidate target point. k =(s1) k +s2 k )·d k , k represents each direction, and the saliency maps of the target candidate points in each direction are merged to obtain the final saliency map:

[0029] Furthermore, targets are extracted based on the final saliency map, specifically including:

[0030] To extract the target using threshold segmentation, first calculate the threshold:

[0031] thresh = mean + 3 / 8 * (max - mean)

[0032] Where mean is the mean of non-zero points on the final saliency map, and max is the maximum value of the final saliency map. Centroids are calculated for points greater than the threshold to obtain the final target position.

[0033] The beneficial effects of this invention are as follows:

[0034] 1. This invention is particularly suitable for detecting weak infrared targets in complex scenes. By screening target candidate points, the false alarm rate can be effectively reduced. By fusing saliency maps in various directions, the background can be effectively suppressed, the target enhanced, and the target detection rate improved.

[0035] 2. This invention is highly robust and not easily affected by noise, background clutter, etc.

[0036] 3. The model used in this invention is simple, has low computational complexity, and fast computation speed, making it suitable for various detection and tracking systems with high real-time requirements. Attached Figure Description

[0037] Figure 1 This is a graph of the first-order derivatives of the Facet model in four directions in this invention;

[0038] Figure 2 This is a flowchart of an infrared weak target detection method based on multi-directional gradient according to the present invention;

[0039] Figure 3 The results of this invention for detecting small infrared targets against a complex sky background are shown in (a)-(f), which are the original image, derivative map, target candidate point map, saliency map in each direction, final saliency map and detection result, respectively. Small targets are marked with boxes on the original image.

[0040] Figure 4 The results of this invention for detecting small infrared targets against a complex background on the sea surface are shown in (a)-(f), which are the original image, derivative map, target candidate point map, saliency map in each direction, final saliency map and detection result, respectively. Small targets are marked with boxes on the original image. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 2 As shown, the present invention provides a method for detecting weak infrared targets based on multi-directional gradients, which specifically includes the following steps:

[0043] The first step is to fit the 5x5 neighborhood image to a polynomial function, as shown below:

[0044] f(r, c) = b0 + b1r + b2c + b3(r) 2 -2)+b4rc+b5(c 2 -2)+b6(r 3 -17r / 5)+b7(r 2 -2)c+b8r(c 2 -2)+b9(c 3 -17c / 5)

[0045] Where r and c are the row and column coordinates within the 5x5 neighborhood image, respectively, and b i (i = 0, 1, ..., 9) are the fitting coefficients, and the first derivative at the horizontal direction angle α is:

[0046]

[0047]

[0048]

[0049] in,

[0050]

[0051] For convolution operation, I is the original image, w i as follows:

[0052]

[0053]

[0054]

[0055]

[0056] The above expression allows us to calculate the derivative graphs (f′0, f′) of the image in four directions (0°, 45°, 90°, 135°). 45 f′ 90 f′ 135 ),like Figure 1 As shown.

[0057] The second step is to find the target location along the directional gradient and determine the target candidate points based on the target locations in each direction.

[0058] In f′0, along the 0-degree direction of the current point (r, c), search to the left for the nearest peak (r) to the current point. max c max And from f′0(r) max c max If f′0(r, c) is monotonically decreasing, obtain the distance dis_left; otherwise, if dis_left = 0, search to the right for the nearest trough (r) to the current point. min c min And from f′0(r) min c min The distance from f′0(r, c) to f′0(r, c) is monotonically increasing, and the distance dis_right is obtained. Otherwise, dis_right = 0. If dis_left > 0, dis_right > 0, and the absolute value of the difference between dis_left and dis_right is within 3, then the current point is considered to be the target point in the 0-degree direction.

[0059] At f′ 45 In the middle, along the 45-degree direction from the current point (r, c), look to the upper left for the nearest peak (r) to the current point. max c max And from f′ 45 (r max c max ) to f′ 45 (r, c) is monotonically decreasing, so we get the x-direction distance dis_left and the y-direction distance dis_up. Otherwise, if dis_left = 0 and dis_up = 0, we search for the nearest trough (r) to the lower right of the current point. min c min And from f′ 45 (r min c min ) to f′ 45(r, c) is monotonically increasing, which gives the distance in the x direction dis_right and the distance in the y direction dis_down. Otherwise, dis_right = 0 and dis_down = 0. If at this time (dis_left + dis_up) > 0 and (dis_right + dis_down) > 0, and the absolute value of the difference between (dis_left + dis_up) and (dis_right + dis_down) is within 3, then the current point is considered to be the target point in the 45-degree direction.

[0060] At f′ 90 In the middle, along the 90-degree direction from the current point (r, c), find the wave peak (r) closest to the current point upwards. max c max And from f′ 90 (r max c max ) to f′ 90 If (r, c) is monotonically decreasing, obtain the y-direction distance dis_up; otherwise, dis_up = 0, and search downwards for the nearest trough (r) to the current point. min c min And from f′ 90 (r min c min ) to f′ 90 (r, c) is monotonically increasing, and the distance in the y direction is obtained as dis_down. Otherwise, dis_down = 0. If dis_up > 0 and dis_down > 0 at this time, and the absolute value of the difference between dis_up and dis_down is within 3, then the current point is considered to be the target point in the 90-degree direction.

[0061] At f′ 135 In the middle, along the 135-degree direction from the current point (r, c), look to the upper right for the nearest peak (r) to the current point. max c max And from f′ 135 (r max c max ) to f′ 135 (r, c) is monotonically decreasing, so we get the x-direction distance dis_right and the y-direction distance dis_up. Otherwise, if dis_right = 0 and dis_up = 0, we search for the nearest trough (r) to the lower left of the current point. min c min And from f′ 135 (r min c min ) to f′ 135(r, c) is monotonically increasing, which gives the distance dis_left in the x direction and the distance dis_down in the y direction. Otherwise, dis_left = 0 and dis_down = 0. If (dis_right + dis_up) > 0 and (dis_left + dis_down) > 0 at this time, and the absolute value of the difference between (dis_right + dis_up) and (dis_left + dis_down) is within 3, then the current point is considered to be the target point in the 135-degree direction.

[0062] If the current point is a target point in the directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees, then the point is considered a candidate target point.

[0063] The third step is to calculate the saliency plot.

[0064] In the four directions (0°, 45°, 90°, 135°), for each target candidate point, based on the peak and trough positions obtained in the previous step, calculate the standard deviation s1 of the peak position within a 3x3 neighborhood, the standard deviation s2 of the trough position within a 3x3 neighborhood, and the difference d between the gradient values ​​at the peak and trough positions in the derivative plot, to obtain the significance plot t. k =(s1) k +s2 k )·d k , k represents each direction, and by merging the saliency maps of each direction, we obtain the final saliency map:

[0065] The fourth step is to extract the target based on the saliency map.

[0066] To extract the target using threshold segmentation, first calculate the threshold:

[0067] thresh = mean + 3 / 8 * (max - mean)

[0068] Where mean is the mean of non-zero points in the saliency plot, max is the maximum value, and the centroid is calculated for points greater than the threshold to obtain the final target position.

[0069] exist Figure 3 In the image, (a) is the original image, (b) is the derivative image in four directions, (c) the white dots are the target candidate points, (d) is the saliency image in four directions, and (e) is the final saliency image after fusion of all directions. Based on the final saliency image, the detection result (f) can be obtained.

[0070] Figure 4 and Figure 3 Similarly.

[0071] The parts of this invention not described in detail are well-known in the art. The embodiments described above are merely preferred embodiments of the present invention, and do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Various modifications and improvements to the technical solutions of this invention made by those skilled in the art without departing from the spirit of the invention should fall within the protection scope defined by the claims of this invention.

Claims

1. A multi-directional gradient-based infrared dim small target detection method, characterized in that, The method comprises: The first-order partial derivatives in each direction neighborhood of the image are calculated by using a Facet model to obtain a direction gradient; target positions in each direction are found according to the direction gradient, information of the target positions in each direction is fused to determine a target candidate point; the target in each direction is enhanced while the background is suppressed to obtain a target candidate point saliency map in each direction, the target candidate point saliency maps in each direction are fused to obtain a final saliency map; and the target is extracted according to the final saliency map; The first-order partial derivatives in a 5x5 neighborhood of the image in each direction are calculated by using a Facet model, wherein the Facet model fits the 5x5 neighborhood image with a polynomial function, and the expression is as follows: where r, c are the row and column coordinates within the 5x5 neighborhood image, respectively, are fitting coefficients, horizontal direction angle The first derivative is: ， wherein , ； The By the following formula: ， wherein, represents a convolution operation, is the original image, is represented as follows: , , The derivative maps of the image in the four directions can be calculated by the above expressions ; The target positions are found according to the direction gradient, and information of the target positions in each direction is fused to determine a target candidate point, and specifically comprising: In , along the 0 degree direction of the current point , find the nearest wave crest to the left of the current point , and if the value from the wave crest to the current point value is monotonically decreasing, then the distance dis_left is obtained, otherwise dis_left=0, find the nearest wave trough to the right of the current point , and if the value from the wave trough value to the current point value is monotonically increasing, then the distance dis_right is obtained, otherwise dis_ right=0, if at this time dis_left>0, dis_right>0, and the absolute value of the difference between dis_left and dis_right is within 3, then the current point is considered to be the target point in the 0 degree direction; exist In the middle, along the current point At a 45-degree angle, look to the upper left for the wave peak closest to the current point. And from the peak value up to the current point value The wave is monotonically decreasing. We obtain the distance in the x-direction (dis_left) and the distance in the y-direction (dis_up). Otherwise, if dis_left = 0 and dis_up = 0, we search for the nearest trough to the current point in the lower right direction. And from the trough value up to the current point value The algorithm is monotonically increasing to obtain the distance in the x-direction dis_right and the distance in the y-direction dis_down. Otherwise, dis_right = 0 and dis_down = 0. If (dis_left + dis_up) > 0 and (dis_right + dis_down) > 0 at this time, and the absolute value of the difference between (dis_left + dis_up) and (dis_right + dis_down) is within 3, then the current point is considered to be the target point in the 45-degree direction. In , along the 90 degree direction of the current point , find the nearest wave crest to the current point upwards , and monotonically decreasing from the wave crest value to the current point value , get the y direction distance dis_up, otherwise dis_up=0, find the nearest wave trough to the current point downwards , and monotonically increasing from the wave trough value to the current point value , get the y direction distance dis_down, otherwise dis_down=0, if dis_up>0 and dis_down>0 at this time, and the absolute value of the difference between dis_up and dis_down is within 3, then the current point is considered to be the target point in the 90 degree direction; exist In the middle, along the current point At a 135-degree angle, look to the upper right for the wave peak closest to the current point. And from the peak value up to the current point value The wave is monotonically decreasing. We obtain the distance in the x-direction (dis_right) and the distance in the y-direction (dis_up). Otherwise, if dis_right = 0 and dis_up = 0, we search for the nearest trough to the lower left of the current point. And from the trough value up to the current point value The algorithm is monotonically increasing to obtain the distance in the x-direction dis_left and the distance in the y-direction dis_down. Otherwise, dis_left=0 and dis_down=0. If (dis_right+dis_up)>0 and (dis_left+dis_down)>0 at this time, and the absolute value of the difference between (dis_right+dis_up) and (dis_left+dis_down) is within 3, then the current point is considered to be the target point in the 135-degree direction. If the current point is a target point in the 0-degree, 45-degree, 90-degree and 135-degree directions, the point is considered as a target candidate point.

2. The multi-directional gradient-based infrared dim small target detection method according to claim 1, characterized in that, The target in each direction is enhanced while the background is suppressed to obtain a target point saliency map in each direction, and the target point saliency maps in each direction are fused to obtain a final saliency map, and specifically comprising: In four directions For each target candidate point, according to the peak and valley positions obtained in the last step, the standard deviation s1 in the 3x3 neighborhood of the peak position, the standard deviation s2 in the 3x3 neighborhood of the valley position, and the difference d between the gradient value of the peak position and the gradient value of the valley position are calculated to obtain the target candidate point saliency map , k represents each direction, and the target candidate point saliency maps of the directions are fused to obtain a final saliency map: .

3. The multi-directional gradient-based infrared dim small target detection method according to claim 1, characterized in that, The target is extracted according to the final saliency map, and specifically comprising: The target is extracted by threshold segmentation, and first, a threshold is calculated: Wherein, mean is the mean value of non-zero points on the final saliency map, max is the maximum value of the final saliency map, the centroid of the points greater than the threshold is calculated to obtain the final target position.

Images