A method for infrared small target detection based on weighted three-layer window local contrast

Abstract

The invention discloses a method for detecting infrared small targets with weighted three-layer window local contrast. First, a layered convolution gradient kernel is used to smooth the complex background. Second, a three-layer window local contrast calculation is performed on the smoothed image to obtain a local contrast map. Then, the complexity of the three-layer window calculation window unit is used to weight the local contrast map. Finally, the target is detected by adaptive threshold segmentation. The invention has good detection performance under complex backgrounds and can significantly enhance the target.

Description

Technical Field

[0001] The invention relates to the field of infrared detection, and in particular to a method for detecting small infrared targets with weighted three-layer window local contrast. Technical Background

[0002] Infrared imaging equipment often has complex background information when detecting targets, and because the target detection distance is far, after imaging by the optical system, the target in the infrared image presents the characteristics of weak contrast between the target and the background and small size. Therefore, small infrared targets do not have shape and texture information, and are often accompanied by complex background clutter and noise in infrared images. Therefore, small targets are easily submerged by complex background clutter, thus placing higher requirements on detection technology.

[0003] Since the multi-frame detection algorithm needs to process a lot of information and the calculation method is relatively complex, it is difficult to meet the real-time requirements in actual engineering applications. Therefore, it is necessary to research and develop a single-frame detection method with good detection performance and strong robustness. In the single-frame detection algorithm, there are generally three mainstream detection methods: filtering-based methods, image data structure-based methods, and visual contrast mechanism-based methods. Typical algorithms based on filtering methods include maximum mean / median filtering methods, mathematical morphology methods, etc.; however, these methods are mainly applicable to scenes with a single, uniform and continuous background, and have poor suppression effects on complex backgrounds. The image data structure-based method regards small targets as sparse components and backgrounds as low-rank components, and transforms the small target detection problem into a process of separating two components from the data structure. Among them, the typical algorithm is the method based on the infrared image block model proposed in 2013. Although this type of method is applicable to various complex scenes, it is complex in calculation and has poor real-time performance. Methods based on visual contrast mechanisms are mainly applicable to scenes where the target has local saliency. After the local contrast calculation method was proposed in 2014, many researchers proposed many improved versions based on the local contrast calculation method, such as adding a preprocessing process to the image or using energy concentration to weight the local contrast. Although these methods have certain improvements in target detection performance, these methods cannot adapt to multi-scale targets due to the fixed window size. Therefore, most of them use multi-scale windows to perform multiple local contrast calculations on the image to adapt to multi-scale targets. However, this approach increases the computational complexity and leads to poor real-time performance of the method. In general, infrared weak target detection technology under complex backgrounds still has problems such as poor robustness, poor complex background clutter suppression, and high false alarm rate. Further research and improvement of the performance of the detection method are still needed. Summary of the invention

[0004] The purpose of the present invention is to overcome the problems of poor detection performance, high false alarm rate, and inability to adapt to multi-scale targets in existing detection technical methods under complex backgrounds. A weighted three-layer window local contrast infrared small target detection method is proposed, which effectively solves the problem that the existing algorithm is sensitive to strong edges and strong noise in complex backgrounds. The method has good complex background suppression effect, significant target enhancement effect, and can handle multi-scale target detection tasks.

[0005] The object of the present invention is achieved through the following solutions:

[0006] A method for detecting small infrared targets by weighted three-layer window local contrast comprises the following steps:

[0007] Step 1: Input a frame of original infrared image;

[0008] Step 2: Use the layered convolution gradient kernel to perform a convolution operation with the original infrared image to suppress the complex background and obtain a smoothed image;

[0009] Step 3: Calculate the local contrast of the image using a three-layer sliding window to obtain a local contrast map;

[0010] Step 4: Set all pixels in the connected domain where the local contrast map has the maximum value to the maximum gray value;

[0011] Step 5: Calculate the window complexity of the original infrared image using a three-layer sliding window to obtain a complexity weighted graph;

[0012] Step 6: Fuse the local contrast map and the complexity weighted map to obtain the target saliency map;

[0013] Step 7: Use adaptive threshold segmentation to segment small targets from the target saliency map and output target information.

[0014] Furthermore, in step 3, the formula used for calculating the local contrast is:

[0015]

[0016]

[0017] I TLCM =D TIB ·D TOB (3)

[0018] Where D TIB is the local contrast between the inner neighborhood unit and the target unit, and the subscript TIB represents the target unit and the inner neighborhood unit; m T is the mean of the target unit, and the subscript T is the target unit area; is the mean value of the inner neighborhood unit of the sliding window, subscript IBi represents the i-th inner neighborhood unit, max() represents the maximum value; D TOB is the local contrast between the target unit and the outer neighborhood unit, and the subscript TOB represents the target unit and the outer neighborhood unit; is the mean value of the inner neighborhood unit of the sliding window, subscript OB j represents the jth outer neighborhood unit, max() represents the maximum value; I TLCM is the local contrast map, and the subscript TLCM represents the local contrast of the three-layer window.

[0019] Furthermore, in step 5, the formula adopted by the complexity weighted graph is:

[0020]

[0021]

[0022]

[0023]

[0024] Where NRIL T is the complexity of the target unit, subscript T is the target unit area; M K is the mean of the first K pixels of the target unit; mean(IB i ) is the mean value of the inner neighborhood unit of the window, IB i is the i-th inner neighborhood unit; is the complexity of the outer neighborhood unit, subscript OB j represents the jth outer neighborhood unit; is the mean of the K largest pixels before the jth outer neighborhood unit in the outer layer of the window, with the subscript OB j represents the jth outer neighborhood unit; is the mean value of the jth neighboring unit in the outer layer of the sliding window, subscript OB j represents the jth outer neighborhood unit; NRIL is the complexity weighted value of the target unit position; I NRIL (x,y) is the complexity weighted value of the pixel (x,y), and the subscript NRIL represents the new local complexity weighting method.

[0025] Furthermore, in step 6, the local contrast map and the complexity weighted map are fused using the following formula:

[0026] I=I TLCM I NRIL (8)

[0027] Where I is the target saliency map, I MTLCMis the local contrast map calculated in step 4. The subscript MTLCM represents the three-layer window local contrast of the maximum connected domain extension. NRIL is the complexity weighted graph obtained in step 5, and the subscript NRIL represents the new local complexity weighting method.

[0028] Furthermore, in step 7, the adaptive threshold segmentation adopts the following formula:

[0029] Th=mean(I)+k·std(I) (9)

[0030] In the formula, Th represents the image segmentation threshold; mean(I) is the mean of the target saliency map, I is the target saliency map, mean() represents the mean of the grayscale values ​​of all pixels in the area; k is the threshold segmentation parameter; std(I) is the standard deviation of the target saliency map, std() represents the standard deviation of the grayscale values ​​of all pixels in the area. When the grayscale value of the pixel in the target saliency map is greater than Th, it is the target pixel.

[0031] The beneficial effects of the present invention are:

[0032] The invention effectively solves the problem that the existing algorithm is sensitive to strong edges and strong noises in complex backgrounds, resulting in low target detection rate, poor strong edge suppression effect and high false alarm rate. The method has good complex background suppression effect and remarkable target enhancement effect. Specifically, the invention utilizes the significance of small infrared targets in infrared images and the gradient information of clutter edges, suppresses strong clutter in complex backgrounds through layered gradient kernel convolution, local contrast calculation, window unit complexity calculation and other means, significantly enhances small targets and effectively reduces the false alarm rate. In addition, the three-layer sliding window adopted by the invention can adapt to the detection of multi-scale targets, reducing the amount of calculation caused by the multi-scale window. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the implementation mode and technical route of the present invention, the following is a brief introduction to the drawings during the implementation of the method:

[0034] Figure 1 The present invention is a flow chart of a method for detecting small infrared targets with weighted three-layer window local contrast.

[0035] Figure 2 Schematic diagram of the layered convolution gradient kernel in step 2 of an embodiment of the present invention.

[0036] Figure 3 Schematic diagram of the three-layer window in step 3 and step 5 of the embodiment of the present invention, where T is the target unit, IB i is the inner neighborhood unit, OBi is the outer neighborhood unit. One unit is 3×3 pixels in size.

[0037] Figure 4 The original infrared image is input in the embodiment of the present invention.

[0038] Figure 5 This is the image after the original infrared image in step 2 of the embodiment of the present invention is convolved with the layered convolution gradient kernel.

[0039] Figure 6 This is a local contrast map in step 3 of the embodiment of the present invention.

[0040] Figure 7 It is the local contrast map in step 4 in the example of the present invention.

[0041] Figure 8 This is the complexity weighted graph of step 5 in the embodiment of the present invention.

[0042] Fig. 9 It is the target saliency map after fusion in step 6 in the embodiment of the present invention.

[0043] Fig.10 This is a small target image obtained after adaptive threshold segmentation in step 7 of the embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0045] As attached Figures 1 to 10 As shown, a method for detecting small infrared targets by weighted three-layer window local contrast includes the following steps:

[0046] Step 1: Input a frame of original infrared image;

[0047] Step 2: Use the layered convolution gradient kernel to perform a convolution operation with the original infrared image to suppress the complex background and obtain a smoothed image;

[0048] Step 3: Calculate the local contrast of the image using a three-layer sliding window to obtain a local contrast map;

[0049] Step 4: Set all pixels in the connected domain where the local contrast map has the maximum value to the maximum gray value;

[0050] Step 5: Calculate the window complexity of the original infrared image using a three-layer sliding window to obtain a complexity weighted graph;

[0051] Step 6: Fuse the local contrast map and the complexity weighted map to obtain the target saliency map;

[0052] Step 7: Use adaptive threshold segmentation to segment small targets from the target saliency map and output target information.

[0053] Furthermore, in step 3, the formula used for calculating the local contrast is:

[0054]

[0055]

[0056] I TLCM =D TIB ·D TOB (3)

[0057] Where D TIB is the local contrast between the inner neighborhood unit and the target unit, and the subscript TIB represents the target unit and the inner neighborhood unit; m T is the mean of the target unit, and the subscript T is the target unit area; is the mean value of the inner neighborhood unit of the sliding window, subscript IB i represents the i-th inner neighborhood unit, max() represents the maximum value; D TOB is the local contrast between the target unit and the outer neighborhood unit, and the subscript TOB represents the target unit and the outer neighborhood unit; is the mean value of the inner neighborhood unit of the sliding window, subscript OB j represents the jth outer neighborhood unit, max() represents the maximum value; I TLCM is the local contrast map, and the subscript TLCM represents the local contrast of the three-layer window.

[0058] Furthermore, in step 5, the formula adopted by the complexity weighted graph is:

[0059]

[0060]

[0061]

[0062]

[0063] Where NRIL T is the complexity of the target unit, subscript T is the target unit area; M K is the mean of the first K pixels of the target unit; mean(IB i ) is the mean value of the inner neighborhood unit of the window, IB i is the i-th inner neighborhood unit; is the complexity of the outer neighborhood unit, subscript OB j represents the jth outer neighborhood unit; is the mean of the K largest pixels before the jth outer neighborhood unit in the outer layer of the window, with the subscript OB j represents the jth outer neighborhood unit; is the mean value of the jth neighboring unit in the outer layer of the sliding window, subscript OBj represents the jth outer neighborhood unit; NRIL is the complexity weighted value of the target unit position; I NRIL (x,y) is the complexity weighted value of the pixel (x,y), and the subscript NRIL represents the new local complexity weighting method.

[0064] Furthermore, in step 6, the local contrast map and the complexity weighted map are fused using the following formula:

[0065] I=I TLCM I NRIL (8)

[0066] Where I is the target saliency map, I MTLCM is the local contrast map calculated in step 4. The subscript MTLCM represents the three-layer window local contrast of the maximum connected domain extension. NRIL is the complexity weighted graph obtained in step 5, and the subscript NRIL represents the new local complexity weighting method.

[0067] Furthermore, in step 7, the adaptive threshold segmentation adopts the following formula:

[0068] Th=mean(I)+k·std(I) (9)

[0069] In the formula, Th represents the image segmentation threshold; mean(I) is the mean of the target saliency map, I is the target saliency map, mean() represents the mean of the grayscale values ​​of all pixels in the area; k is the threshold segmentation parameter; std(I) is the standard deviation of the target saliency map, std() represents the standard deviation of the grayscale values ​​of all pixels in the area. When the grayscale value of the pixel in the target saliency map is greater than Th, it is the target pixel.

[0070] In other embodiments of the present invention, the following steps are included: first, the original infrared image is input; second, the complex background is smoothed using a layered convolution gradient kernel; then, a three-layer window local contrast calculation is performed on the smoothed image to obtain a local contrast map; third, the local contrast map is weighted using the complexity of the three-layer window calculation window unit; finally, the target is detected through adaptive threshold segmentation and the target position information is output.

Claims

1. A method for detecting small infrared targets based on weighted three-layer window local contrast, characterized in that: Includes steps: Step 1: Input a frame of original infrared image; Step 2: Use the layered convolution gradient kernel to perform a convolution operation with the original infrared image to suppress the complex background and obtain a smoothed image; Step 3: Calculate the local contrast of the image using a three-layer sliding window to obtain a local contrast map; Step 4: Set all pixels in the connected domain where the local contrast map has the maximum value to the maximum gray value; Step 5: Calculate the window complexity of the original infrared image using a three-layer sliding window to obtain a complexity weighted graph; Step 6: Fuse the local contrast map and the complexity weighted map to obtain the target saliency map; Step 7: Use adaptive threshold segmentation to segment small targets from the target saliency map and output target information; The formula adopted by the complexity weighted graph described in step 5 is: Where NRIL T is the complexity of the target unit, subscript T is the target unit area; M K is the mean of the first K pixels of the target unit; mean(IB i ) is the mean of the inner neighborhood units of the window, IB i is the i-th inner neighborhood unit; is the complexity of the outer neighborhood unit, subscript OB j represents the jth outer neighborhood unit; is the mean of the K largest pixels before the jth outer neighborhood unit in the outer layer of the window, with the subscript OB j represents the jth outer neighborhood unit; is the mean value of the jth neighborhood unit in the outer layer of the sliding window, with the subscript OD j represents the jth outer neighborhood unit; NRIL is the complexity weighted value of the target unit position; I NRIL (x,y) is the complexity weighted value of the pixel (x,y), and the subscript NRIL represents the new local complexity weighting method.

2. The infrared small target detection method based on weighted three-layer window local contrast according to claim 1 is characterized in that: The method for suppressing complex background described in step 2 is to perform filtering operation using a layered convolution gradient kernel.

3. The infrared small target detection method based on weighted three-layer window local contrast according to claim 1 is characterized in that: The three-layer sliding window described in step 3 calculates the local contrast, and the calculation formula is: I TLCM =D TIB ·D TOB (3) Where D TIB is the local contrast between the inner neighborhood unit and the target unit, and the subscript TIB represents the target unit and the inner neighborhood unit; m T is the mean of the target unit, and the subscript T is the target unit area; is the mean value of the inner neighborhood unit of the sliding window, subscript IB i represents the i-th inner neighborhood unit, max() represents the maximum value; D TOB is the local contrast between the target unit and the outer neighborhood unit, and the subscript TOB represents the target unit and the outer neighborhood unit; is the average value of the outer neighborhood unit of the sliding window, subscript OB j represents the jth outer neighborhood unit, max() represents the maximum value; I TLCM is the local contrast map, and the subscript TLCM represents the three-layer window local contrast.

4. The infrared small target detection method based on weighted three-layer window local contrast according to claim 1 is characterized in that: The local contrast map and the complexity weighted map fusion described in step 6 adopt the following formula: I=I MTLCM ·I NRIL (8) Where I is the target saliency map, I MTLCM is the local contrast map calculated in step 4. The subscript MTLCM represents the three-layer window local contrast of the maximum connected domain extension. NRIL is the complexity weighted graph obtained in step 5. The subscript NRIL represents the new local complexity weighting method.

5. The infrared small target detection method based on weighted three-layer window local contrast according to claim 1 is characterized in that: The adaptive threshold segmentation in step 7 adopts the following formula: Th=mean(I)+k·std(I) (9) Where Th represents the image segmentation threshold; mean(I) is the mean of the target saliency map, I is the target saliency map, mean() represents the mean of the grayscale values ​​of all pixels in the area; k is the threshold segmentation parameter; std(I) is the standard deviation of the target saliency map, std() represents the standard deviation of the grayscale values ​​of all pixels in the area; when the grayscale value of the pixel in the target saliency map is greater than Th, it is the target pixel.

Images