A weak signal detection method based on motion energy estimation

Abstract

The application aims to provide a weak signal detection method based on motion energy estimation, comprising the following steps: A, constructing a motion model of 12 target signals in an image, and the motion model has a size of 5*5; B, multiplying the sequence image with the motion model to obtain an image to be detected; C, scanning all pixels in the current frame image one by one, taking the neighborhood of the currently scanned pixel as the center to construct the center neighborhood block and the corresponding neighborhood blocks in the four directions above, below, left and right, calculating the gradient, and taking the position of the pixel point as a candidate target position if the gradient in more than three of the four directions is greater than a predetermined gradient threshold; D, performing high-order energy correlation enhancement on the candidate target point, obtaining the energy maximum value of the target, and outputting a target image; E, performing the same processing on the sequence image, superimposing all target images, and obtaining a target motion trajectory image. The application can effectively eliminate noise and extract a real target, and greatly improve the detection rate of the algorithm.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer image processing, in particular to a weak signal detection method based on motion energy estimation. BACKGROUND

[0002] Infrared weak target detection is widely used in military, remote sensing monitoring, forest fire monitoring and other aspects. However, due to the influence of atmospheric aerosol, long detection distance, random noise generated in the imaging process and other factors, the target appears as a small bright spot on the image, lacks texture features, and even is submerged in a complex background, which brings great difficulty to small target detection.

[0003] In the detection of weak targets under low signal-to-noise ratio, the target energy may be lower than the clutter energy, and after image preprocessing, it is still difficult to extract the target by threshold segmentation, and single-frame detection and multi-frame correlation detection need to be combined. The traditional pipeline filtering is one of the common multi-frame detection algorithms, but the pipeline filtering has great limitations. When the target does not move or the background moves quickly, the algorithm fails. SUMMARY

[0004] The present application aims to provide a weak signal detection method based on motion energy estimation, which can effectively eliminate noise and extract real targets, greatly improving the detection rate of the algorithm.

[0005] The technical scheme of the present application is as follows:

[0006] The weak signal detection method based on motion energy estimation comprises the following steps:

[0007] A. Construct 12 kinds of target signal motion models in the image, and the motion model size is 5*5; wherein the motion directions of the first to fourth motion models are horizontal, vertical and two diagonal directions, and the motion directions of the fifth to twelfth motion models are random directions;

[0008] B. Multiply the sequence image by the motion model to obtain the image to be detected;

[0009] C. Set the size of the neighborhood block and the gradient threshold, and scan all the pixels in the current frame image one by one, construct a center neighborhood block with the neighborhood of the currently scanned pixel as the center, then select neighborhood blocks with the same size in the upper, lower, left and right directions of the center neighborhood block, calculate the gray mean value, and calculate the gradient between the center neighborhood block and the neighborhood blocks in the four directions. If the gradient of more than three directions is greater than the pre-set gradient threshold, the pixel is considered as a candidate target point, otherwise it is eliminated;

[0010] D. Perform high-order correlation energy enhancement on the candidate target points, find the energy maximum value of the target points, and output the target image.

[0011] E, the same processing is carried out on the sequence image, and target images of each frame image are obtained, and all target images are superimposed to obtain a target motion trajectory image.

[0012] In the step B, the sequence image is preprocessed by anisotropic background estimation, and then multiplied by a motion model.

[0013] The preprocessed image is multiplied by a motion model, and the expression is:

[0014]

[0015] Wherein, k=1, 2,..., 12, represents the serial number of 12 kinds of motion models, x, y represents the position of the pixel, m, n represents the radius of the template, T k (x, y) represents the motion model, F0 represents the preprocessed image, t N represents the Nth frame image.

[0016] The size of the neighborhood block is 3*3-11*11.

[0017] In the step C, the formula for constructing the center neighborhood block and the corresponding neighborhood blocks in four directions and calculating the mean value is:

[0018]

[0019] Wherein, mid represents the target neighborhood; up_mean, down_mean, left_mean, right_mean, mid_mean respectively represent the mean values of the upper, lower, left, right and center neighborhood blocks, and L is the size of the neighborhood block.

[0020] In the step C, the gradient between the center neighborhood block and the corresponding four direction neighborhood blocks is calculated, and the formula for sorting the gradients in four directions is:

[0021]

[0022] Wherein, data1, data2, data3, data4 are the absolute values of the gradient difference between the center neighborhood and the four direction neighborhood; A is the gradient difference between the current pixel neighborhood and the four direction neighborhood.

[0023] The step D is to enhance the high-order correlation energy of the candidate target point and obtain the target energy maximum value, and the specific method is as follows: taking the position of the candidate target pixel in the current frame image, finding the neighborhood of the position in the next frame image, multiplying the neighborhood with the pixel value of the candidate target in the current frame and summing, multiplying the sum with the neighborhood of the position of the candidate target point in the third frame image and summing, dividing the obtained result by the neighborhood block size, and replacing the candidate target pixel value of the current frame image with the operation result.

[0024] The specific expression is as follows:

[0025]

[0026] Wherein, target1, target2, target3 represent three adjacent frames of images, target1 represents the position of the target at time n, target2 represents the neighborhood of the position of the target at time n+1, target3 represents the neighborhood of the position of the target at time n+2, r is the search neighborhood radius of the target, and L is the size of the neighborhood block. 1st_min The minimum value of the four direction gradient differences in A sorted from small to large, data 2nd_min The second minimum value of the four direction gradient differences in A sorted from small to large, T is a preset gradient threshold, and E k1 (i,j,t n ) is the high-order correlation of the previous two frames of images, E2(i,j,t n ) is the result of the high-order correlation of the previous two frames of images again with the high-order correlation of the third frame of image, F0'(x,y,t N ) is the target energy maximum value.

[0027] The formula for superimposing all target images in the step E is as follows:

[0028] E k =F'0(x,y,t1)+F'0(x,y,t2)+...+F'0(x,y,t N ) ⑸

[0029] E k is the result of superimposing N frames of target images, that is, the target motion trajectory.

[0030] Compared with the prior art, the present application has the following advantages:

[0031] The present application can enhance the continuous motion energy of the target along the motion direction, and the occurrence of noise is often random. The image preprocessed image and the motion model are energy fused, the discrimination model of the salient region is combined, the purpose of removing noise and extracting the real target is achieved, and the detection rate of the algorithm is greatly improved.

[0032] The application achieves the purpose of suppressing noise by gradient difference between data blocks through high-order correlation saliency discrimination model, so that the algorithm still has good detection performance in non-stationary area. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The 12 motion model graphs for the structure of example 1;

[0034] Figure 2 The structural graph for the center neighborhood block example of example 1;

[0035] Figure 3 The original image and the motion estimation result graph of the first frame of scene 1 of example 1;

[0036] Figure 4 The original image and the motion estimation result graph of the 30th frame of scene 2 of example 1;

[0037] Figure 5 The original image and the motion estimation result graph of the first frame of scene 3 of example 1;

[0038] Figure 6 The original image and the motion estimation result graph of the first frame of scene 4 of example 1;

[0039] Figure 7 Comparison of motion trajectories obtained by different algorithms for scene 1;

[0040] Figure 8 Comparison of motion trajectories obtained by different algorithms for scene 2;

[0041] Figure 9 Comparison of motion trajectories obtained by different algorithms for scene 3;

[0042] Figure 10 Comparison of motion trajectories obtained by different algorithms for scene 4. DETAILED DESCRIPTION

[0043] The application will be specifically described below in combination with the drawings and examples.

[0044] Example 1

[0045] The weak signal detection method based on motion energy estimation of the present embodiment comprises the following steps:

[0046] The weak signal detection method based on motion energy estimation comprises the following steps:

[0047] A, construct 12 kinds of target signals in the image motion model, the motion model size is 5*5, such as Figure 1The motion directions of the first to fourth motion models are horizontal, vertical and two diagonal directions respectively, and the motion directions of the fifth to twelfth motion models are random directions.

[0048] B. The preprocessed image is multiplied by the motion model to obtain the image to be detected; the preprocessing is specifically filtering by anisotropic background estimation to remove clutter in the image;

[0049] The preprocessed image is multiplied by the motion model, and the expression is:

[0050]

[0051] Wherein, k = 1, 2,..., 12, represents the serial number of the 12 motion models, x, y represents the position of the pixel, m, n represents the radius of the template, T k (x, y) represents the motion model, F0 represents the preprocessed image, t N represents the Nth frame image.

[0052] C. The size of the neighborhood block and the gradient threshold are set, and all pixels in the current frame image are scanned one by one. The center neighborhood block is constructed with the neighborhood of the currently scanned pixel as the center. Then the neighborhood blocks with the same size are selected in the upper, lower, left and right directions of the center neighborhood block, and the gray mean values are calculated. The gradients between the center neighborhood block and the four direction neighborhood blocks are calculated respectively. If the gradients of more than three directions are greater than the pre-set gradient threshold, the pixel is considered as a candidate target point, otherwise it is removed.

[0053] In this embodiment, the constructed neighborhood block is a 3*3 block containing nine pixels; an example of the neighborhood block is shown in the following table: Figure 2 As shown in the figure, 204 is the center pixel, and the pixels contained in the center neighborhood block are 204 and the surrounding green blocks, that is, in the figure: 62 69 71 71 204 61 71 68 61

[0057] The formula for constructing the center neighborhood block and the corresponding four direction neighborhood blocks and calculating the mean values is:

[0058]

[0059] Wherein, mid represents the neighborhood of the target; up_mean, down_mean, left_mean, right_mean and mid_mean represent the mean values of the upper, lower, left, right and center neighborhood blocks respectively, and L is the size of the neighborhood block.

[0060] The formula for calculating the gradients between the central neighborhood block and the corresponding four directional neighborhood blocks, and then sorting the gradients in the four directions, is as follows:

[0061]

[0062] Where data1, data2, data3, and data4 are the absolute values ​​of the gradient differences between the central neighborhood and the four directional neighborhoods; A is the gradient difference between the current pixel's neighborhood and the four directional neighborhoods.

[0063] D. Perform high-order correlation energy enhancement on the candidate target points, find the energy maxima of the target points, and output the target image;

[0064] The method for performing higher-order correlation energy enhancement on candidate target points and finding the target energy maxima is as follows:

[0065] Find the location of the candidate target pixel in the current frame image, find the neighborhood of that location in the next frame image, multiply the neighborhood with the pixel value of the candidate target in the current frame and sum them, multiply the sum with the neighborhood of the candidate target point in the third frame image and sum them, divide the result by the size of the neighborhood block, and replace the candidate target pixel value in the current frame image with the result of the operation.

[0066] The specific expression is as follows:

[0067]

[0068] Where target1, target2, and target3 represent three adjacent frames, target1 represents the target's location at time n, target2 represents the neighborhood of the target's location at time n+1, target3 represents the neighborhood of the target's location at time n+2, r is the radius of the target's search neighborhood, and L is the size of the neighborhood block. 1st_min This represents the minimum value of the gradient differences in the four directions of A, sorted from smallest to largest. 2nd_min This represents the second smallest value of the gradient differences in the four directions of A, sorted from smallest to largest. T is the preset gradient threshold, and E is the second smallest value. k1 (i,j,t n E2(i,j,t) represents the higher-order correlation between the first two frames. n The result of the high-order correlation between the first two frames is then correlated with the high-order correlation of the third frame, F'0(x,y,t) N ) represents the target energy maximum.

[0069] E. Perform the same processing on the sequence images to obtain the target images of each frame. Superimpose all the target images to obtain the target motion trajectory image.

[0070] In step E, the formula for superimposing all target images is:

[0071] E k =F'0(x,y,t1)+F'0(x,y,t2)+...+F'0(x,y,t N ) ⑸.

[0072] This invention Figures 3-6 Here are some examples of image processing results for some frames: Figure 3 The left image is the original image of the first frame of Scene 1, and the right image is the motion estimation result obtained by the method of this embodiment for the first frame of Scene 1. Figure 4 The left image is the original image of frame 30 in scene 2, and the right image is the motion estimation result obtained by the method of this embodiment in frame 30 of scene 2. Figure 5 The left image in the middle is the original image of the first frame of scene 3, and the right image is the motion estimation result obtained by the method of this embodiment for the first frame of scene 3. Figure 6 The left image is the original image of the first frame of scene 4, and the right image is the motion estimation result obtained by the method of this embodiment for the first frame of scene 4.

[0073] Example 2

[0074] To verify the effectiveness of the detection algorithm of this invention, four scenarios were selected for experimentation and compared with traditional algorithms. These comparison algorithms are Top-hat, WLDM, MPCM, MGDWE, TLLCM, and TDLMS. The algorithm of this invention is named Proposed in the figure. The parameters used in the algorithm of this invention are shown in Table 1 below, where M is the cumulative number of frames, L is the size of the spatial neighborhood of the target, and r is the radius of the target search neighborhood, referring to formula (3).

[0075] Table 1. Parameters related to motion direction estimation and multi-frame correlation detection

[0076] Frame number M Spatial neighborhood L Target search neighborhood radius r Features Scenario 1 3 5 9*9 Target and background move slowly Scenario 2 3 5 9*9 Target moves slowly, background almost not moving Scenario 3 3 11 19*19 Target moves faster, background changes slowly Scenario 4 3 3 7*7 Target moves fast, background span large

[0077] Experimental results are as follows Figures 7-10 As shown:

[0078] Figure 7 To compare the motion trajectories obtained by different algorithms in Scene 1, from Figure 7 As can be seen, in scenario 1, the target motion trajectory extracted by the WLDM algorithm is covered by the background motion trajectory. The target motion trajectory extracted by the MPCM algorithm is discontinuous. The MGDWE algorithm, TLLCM algorithm, improved TOP-HAT and TDLMS algorithms can extract the target motion trajectory, but all have varying degrees of noise interference.

[0079] Figure 8The motion trajectory comparison obtained by different algorithms in scene 2 is shown in Fig. 2. Figure 8 As shown in Fig. 2, in scene 2, all the algorithms can extract the motion trajectory of the target completely, but there is still a small amount of background clutter interference.

[0080] Figure 9 The motion trajectory comparison obtained by different algorithms in scene 3 is shown in Fig. 3. Figure 9 As shown in Fig. 3, in scene 3, the motion trajectory extracted by WLDM, MPCM and MGDWE algorithms cannot be observed, and the noise interference is serious. The TLLCM algorithm, TOP-HAT and TDLMS algorithm can extract the motion trajectory of the target, but cannot avoid the clutter interference.

[0081] Figure 10 The motion trajectory comparison obtained by different algorithms in scene 4 is shown in Fig. 4. Figure 10 As shown in Fig. 4, in scene 4, the trajectory obtained by the WLDM algorithm is submerged in the serious background clutter, and the MPCM algorithm and the MGDWE algorithm can observe the trajectory, but the background moving trajectory on the image is obvious. The TLLCM algorithm, TOP-HAT and TDLMS algorithm can better extract the motion trajectory, but there is different degree of noise interference.

[0082] Compared with the traditional detection algorithm, the detection algorithm (Proposed) proposed in the present application can more effectively suppress the background clutter and completely extract the motion trajectory of the target in the above four scenes.

[0083] From the above analysis, it can be seen that the WLDM and MGDWE algorithms mainly use the local difference method to detect the target in the image, which can easily lead to a high false detection rate and a high missing detection rate. The MPCM algorithm detects the target by using the local contrast information of the multi-scale block, but the algorithm easily identifies the dark background as the target. Although the TLLCM algorithm uses a three-layer filtering window, the generalization ability of the algorithm is poor. Therefore, the local contrast measurement algorithm is suitable for a stable background with high contrast, but is not suitable for a non-stationary and dramatic change scene. This is because the local contrast measurement algorithm mainly uses the local features of the target to detect the weak target, and therefore has a high requirement for the contrast of the background. However, the real background is often complex and changeable. The TOP-HAT algorithm seriously depends on the selection of the structure element, and the TDLMS algorithm depends on the iteration step. Therefore, the traditional filtering algorithm is suitable for a target with small size and a uniform and stable background, but the detection effect is greatly reduced in a scene with complex background components. The algorithm of the present application is not only suitable for a stable background scene, but also can achieve good detection effect in a large-span and dynamic change scene.

Claims

1. A weak signal detection method based on motion energy estimation, characterized in that... Includes the following steps: A. Construct motion models of 12 target signals in the image, with each motion model having a size of 5*5; The motion directions of motion models 1-4 are horizontal, vertical, and two diagonal directions, respectively, while the motion directions of motion models 5-12 are random. B. Multiply the sequence images by the motion model to obtain the image to be detected; The preprocessed images are multiplied by the motion model, and the expression is: ⑴ in, , representing the sequence number of the 12 motion models, x, y represent the pixel position, m, n represent the radius of the template, Representative motion model, This represents the preprocessed image. Represents the Nth frame image; C. Set the neighborhood block size and gradient threshold, scan all pixels in the current frame image one by one, construct a central neighborhood block with the neighborhood of the currently scanned pixel as the center, then select neighborhood blocks of the same size in the four directions of the central neighborhood block, calculate their gray value, and calculate the gradient between the central neighborhood block and the neighborhood blocks in the four directions respectively. If the gradient in three or more of the four directions is greater than the preset gradient threshold, the pixel is considered a candidate target point; otherwise, it is discarded. D. Perform high-order correlation energy enhancement on the candidate target points, find the energy maxima of the target points, and output the target image; The method for performing high-order correlation energy enhancement on candidate target points and finding the target energy maxima is as follows: take the position of the candidate target pixel in the current frame image, find the neighborhood of that position in the next frame image, multiply the neighborhood with the pixel value of the candidate target in the current frame and sum them, multiply the sum with the neighborhood of the candidate target point in the third frame image and sum them, divide the result by the neighborhood block size, and replace the candidate target pixel value in the current frame image with the result. The specific expression is as follows: ⑷ in, target1, target2, target3 Represents three adjacent frames target1 This represents the location of the target at time n. target2 This represents the neighborhood of the target's location at time n+1. target3 This represents the neighborhood of the target's location at time n+2. The search neighborhood radius of the target. The size of the neighboring block. Representative to A The minimum value of the gradient differences in the four directions, sorted from smallest to largest. Representative to A The second smallest value after sorting the gradient differences in the four directions from smallest to largest. T To preset the gradient threshold, The first two frames are related by higher order. The results of the higher-order correlation of the first two frames are then correlated with the higher-order correlation of the third frame. The target energy is the maximum value; E. Perform the same processing on the sequence of images to obtain the target images of each frame. Superimpose all the target images to obtain the target motion trajectory image.

2. The weak signal detection method based on motion energy estimation as described in claim 1, characterized in that: The preprocessing in step B specifically involves filtering through anisotropic background estimation to remove background clutter from the image. In step B, the sequence images are first preprocessed by anisotropic background estimation, and then multiplied with the motion model.

3. The weak signal detection method based on motion energy estimation as described in claim 1, characterized in that: The size of the neighboring block is 3*3 to 11*11.

4. The weak signal detection method based on motion energy estimation as described in claim 1, characterized in that: In step C, the formula for constructing the central neighborhood block and its corresponding four directional neighborhood blocks and calculating the average is as follows: ⑵ in, mid Represents the neighborhood where the target is located; up_mean , down_mean , left_mean, right_mean , mid_mean These represent the average values ​​for the top, bottom, left, right, and center neighboring blocks, respectively. L The size of the neighboring block.

5. The weak signal detection method based on motion energy estimation as described in claim 4, characterized in that: In step C, the formula for calculating the gradient between the central neighborhood block and the corresponding four directional neighborhood blocks, and sorting the gradients in the four directions, is as follows: ⑶ in, data1 , data2 , data3 , data4 It is the absolute value of the gradient difference between the central neighborhood and the four directional neighborhoods; A This represents the gradient difference between the current pixel's neighborhood and its four directional neighborhoods.

6. The weak signal detection method based on motion energy estimation as described in claim 1, characterized in that: In step E, the formula for superimposing all target images is: ⑸ This is the result of overlaying N frames of target images, i.e., the target's motion trajectory.

Images