A method for infrared small target detection based on edge window guided filtering

Abstract

The invention discloses an infrared small target detection method based on edge window guided filtering, and the steps are as follows: firstly, a fast saliency filtering method is used to obtain a priori image of the infrared small target; then, eight different hollow structure edge window filters are used to guide filtering of the original image, and the saliency prior is added as a weight term to the guided filtering process, and the calculated filtering result is a background estimation image; the original image and the background estimation image are subtracted to obtain a target image, and a threshold segmentation method is used to output the target detection result. The invention proposes a hollow structure edge window guided filter for the first time, and uses it for background estimation and detection of infrared small targets, with a simple implementation process and good real-time performance, and can effectively improve the probability of small target detection under complex backgrounds and reduce the probability of false alarms caused by edge clutter.

Description

Technical Field

[0001] The invention belongs to the field of infrared image processing and small target detection, and in particular relates to an infrared small target detection method based on edge window guided filtering. Background Art

[0002] Infrared detection methods have been widely used in military fields such as target early warning search and civilian fields such as field rescue and rescue, and have achieved remarkable results. In various application fields, real-time infrared small target detection technology with high detection rate and low false alarm rate has become an inevitable requirement for practical engineering applications. However, due to the long distance between the target to be detected and the detector, the imaging area of ​​the infrared small target is very small, and the background area is often interfered by clutter such as trees, clouds, waves, and buildings. The above factors have brought many difficulties to the accurate detection of infrared small targets.

[0003] At present, many scholars have proposed infrared small target detection methods based on single frames from different perspectives because they are relatively simple to implement and can be used as a basic link for multi-frame sequence target track association. Commonly used single-frame detection methods include background estimation methods, human visual system methods, low-rank sparse decomposition methods, deep learning methods, etc. Background estimation methods usually use spatial or transform domain filters to analyze and estimate the background, which usually has low computational complexity and is simple to implement, but the detection results often contain edges or corners with obvious features; the human visual system uses the local contrast mechanism and takes the feature dissimilarity between the target center position and the surrounding neighborhood position as the measurement benchmark. It can enhance the target while suppressing the background, but heterogeneous clutter background with similar target significant characteristics will cause false alarms and false detections; low-rank sparse decomposition methods use the assumptions of low-rank background and sparsity of targets to transform the target detection problem into an optimization problem, but have disadvantages such as high computational complexity and difficulty in engineering; deep learning methods can automatically obtain multi-level and multi-dimensional feature information from images, and a variety of algorithms have emerged, such as small target detection methods based on generative adversarial networks, attention-based local contrast networks, and cross-connected bidirectional pyramid networks. However, the above methods are difficult to cope with actual situations such as missing large data samples or changes in application scenarios.

[0004] In summary, the current infrared small target detection algorithms have shortcomings such as low robustness, poor real-time performance, and weak anti-interference ability to varying degrees. It is still necessary to seek a real-time small target detection method with high detection rate and low false alarm rate in various complex scenes. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an infrared small target detection method based on side window guided filtering, which can effectively suppress false alarms caused by complex backgrounds, and has the characteristics of strong anti-interference ability and low algorithm complexity, and can meet the actual application needs of engineering.

[0006] The technical solution of the present invention is: a method for detecting small infrared targets based on edge window guided filtering, the method comprising the following steps:

[0007] Step 1.1, obtain the original infrared image D, use the fast saliency filtering method to calculate the infrared small target prior image C;

[0008] Step 1.2, eight different hollow-structured edge window filters are used to perform weighted guided filtering on the original infrared image D to obtain a background estimation image Q;

[0009] Step 1.3: Subtract the original infrared image D from the background estimation image Q to obtain the target image T, use the threshold segmentation method to determine the position of the small infrared target, and output the target detection result.

[0010] Furthermore, in the fast saliency filtering method described in step 1.1, the filter is a spatial domain filter kernel of size 5×5;

[0011] Furthermore, the eight different hollow-structured side window filters described in step 1.2 have effective filtering ranges of (2R out +1)×(2R out +1) The left half, right half, upper half, lower half, lower left 1 / 4, lower right 1 / 4, upper right 1 / 4, upper left 1 / 4 of the square area, R out is the outer diameter of the filter, indicating the radius size of the target area and the background area. The target pixel is located at the edge center or corner point of the effective filtering area. Based on the above effective filtering area, the size of (2R in +1)×(2R in +1), R in is the inner diameter of the filter, indicating the radius of the target area only. The filter presents a hollow structure with a side window shape (see Figure 2 );

[0012] Furthermore, in the weighted guided filtering described in step 1.2, the weight term of each pixel is divided into w k and w′ k Two parts, the calculation formula is: w k =c k λ, where c k is the pixel saliency calculated in step 1.1, cmean is the average value of the whole image of the saliency prior map C, c min is the minimum value of the saliency prior map C, and λ is the regularization parameter;

[0013] Furthermore, the threshold segmentation described in step 1.3: the product of the maximum value of the target image T and the constant coefficient m is used as the segmentation threshold, and the value range of m is 0.3 to 0.5.

[0014] The advantages of the present invention compared with the prior art are:

[0015] (1) Using visual saliency as a priori weight term to guide filtering is beneficial to improving the distinction between small infrared targets and background areas, and better completing background preservation and target smoothing;

[0016] (2) The hollow-structured edge window-guided filter can maintain various edges and corners while smoothing the target, reducing the residual clutter in the background estimation image and lowering the false alarm probability of the algorithm;

[0017] (3) Both the fast saliency filtering method and the edge window guided filtering method are simple to implement and have good parallelism, which can meet the real-time requirements of the project. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flow chart of an infrared small target detection method based on edge window guided filtering of the present invention;

[0019] Figure 2 Eight different hollow structure side window filter shapes proposed by the present invention;

[0020] Figure 3 An infrared image containing a small target is exemplified in the present invention;

[0021] Figure 4 The present invention is Figure 3 Constructed target saliency prior map;

[0022] Figure 5 The present invention is Figure 3 and Figure 4 The background image and target image are obtained by edge window weighted guided filtering, where: Figure 5 (a) is the background image, Figure 5 (b) is the target image;

[0023] Figure 6 The present invention is Figure 5 The detection result of the target image in is obtained by threshold segmentation. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The present invention adopts a hollow structure side window filter for the first time, and adds the saliency prior information of the small target as a weight term to the guided filtering process, effectively estimates the background image, reduces the influence of clutter such as edges and corners on the detection results, and realizes fast and accurate detection of infrared small targets.

[0025] like Figure 1 As shown, a method for detecting small infrared targets based on edge window guided filtering is described in detail below, including the following steps:

[0026] 1. Get the original infrared image D, such as Figure 3 shown.

[0027] 2. Calculate the saliency prior map C of the infrared small target, such as Figure 4 The specific steps include:

[0028] 2.1. Use a 2×2 mean filter to smooth the image.

[0029] 2.2. The image after mean filtering is subjected to saliency filtering using a 5×5 spatial filter kernel F to obtain a priori image C. The spatial filter kernel F is derived based on the facet model (see A fast-saliency method for real-time infrared small target detection. Infrared Physics & Technology [J]. 2016, 77: 440-450), and its form is as follows:

[0030]

[0031] 3. Based on the original side window guided filtering method (see Side window guided filtering. Signal Processing [J]. 2019, 165: 315-330), eight different hollow structure side window filters are further constructed, and the saliency prior map C calculated in step 2 is used to complete the weighted guided filtering of the original image D. The specific steps include:

[0032] 3.1. Construct eight different hollow-structured side window filters, whose effective filtering ranges are (2R out +1)×(2R out +1) The left half, right half, upper half, lower half, lower left 1 / 4, lower right 1 / 4, upper right 1 / 4, upper left 1 / 4 of the square area, R outis the outer diameter of the filter, indicating the radius size of the target area and the background area. The target pixel is located at the edge center or corner point of the effective filtering area. Based on the above effective filtering area, the size of (2R in +1)×(2R in +1), R in is the inner diameter of the filter, indicating the radius of the target area only. The filter presents a hollow structure with a side window shape (see Figure 2 );

[0033] 3.2. Calculate the prior weight w of each pixel k and w′ k , the calculation formula is:

[0034] w k =c k λ

[0035]

[0036]

[0037] Among them, c k is the pixel saliency calculated in step 2, c mean is the average value of the whole image of the saliency prior map C, c min is the minimum value of the saliency prior map C, and λ is the regularization parameter;

[0038] 3.3, according to the eight different hollow structure edge window filters described in step 3.1, calculate each pixel point in the effective area The mean within and variance And get the linear coefficient and The calculation formula is as follows:

[0039]

[0040]

[0041] Among them, w k and w′ k is the prior weight calculated in step 3.2. The superscript n in the formula represents eight different shapes of edge window filters n = {1, 2, 3, 4, 5, 6, 7, 8};

[0042] 3.4. Linear coefficients for each pixel position and Average the effective area of ​​the corresponding edge window filter to get the average value and Then use the gray value d of the corresponding position of the original image i , get the gray value of the filtered image The calculation formula is as follows:

[0043]

[0044]

[0045]

[0046] in, is the number of pixels in the effective area of ​​the edge window filter;

[0047] 3.5. Select the grayscale d of the original image from the eight edge window guided filtering results. i The closest value is used as the output of the background estimation image Q, and the calculation formula is as follows:

[0048]

[0049] 4. Subtract the original infrared image D from the background estimation image Q to obtain the target image T, as Figure 5 shown. Figure 5 The present invention is Figure 3 and Figure 4 The background image and target image are obtained by edge window weighted guided filtering. Figure 5 (a) is the background image, Figure 5 (b) is the target image.

[0050] 5. Perform threshold segmentation on the target image T, determine the position of the infrared small target, and output the target detection result, such as Figure 6 As shown, the segmentation threshold is the product of the maximum value of the target image T and the constant coefficient m, and the value range of m is 0.3 to 0.5.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. A method for detecting small infrared targets based on edge window guided filtering, characterized in that: The method comprises the following steps: Step 1.1, obtain the original infrared image D, use the fast saliency filtering method to calculate the infrared small target prior image C; Step 1.2, eight different hollow-structured edge window filters are used to perform weighted guided filtering on the original infrared image D to obtain a background estimation image Q; Step 1.3, subtract the original infrared image D from the background estimation image Q to obtain the target image T, use the threshold segmentation method to determine the position of the infrared small target, and output the target detection result; Weighted guided filtering as described in step 1.2: The weight term of each pixel is divided into and Two parts, the calculation formula is: , , ,in is the pixel saliency size calculated in step 1.1, is the average value of the whole image of the significant prior map C, is the minimum value of the saliency prior map C. is the regularization parameter; According to the eight different hollow structure edge window filters, each pixel point in the effective area is calculated. The mean within and variance , and obtain the linear coefficient and , the calculation formula is as follows: in, and The superscript n in the formula represents eight different shapes of edge window filters. .

2. The infrared small target detection method based on edge window guided filtering according to claim 1 is characterized in that: The fast saliency filtering method described in step 1.1: the filter is a spatial filter kernel of size 5×5.

3. The infrared small target detection method based on edge window guided filtering according to claim 1 is characterized in that: The eight different hollow-structure side window filters described in step 1.2 have effective filtering ranges of (2R out +1)×(2R out +1) The left half, right half, upper half, lower half, lower left 1 / 4, lower right 1 / 4, upper right 1 / 4, upper left 1 / 4 of the square area, R out is the outer diameter of the filter, indicating the radius size of the target area and the background area. The target pixel is located at the edge center or corner point of the effective filtering area. Based on the above effective filtering area, the size of (2R in +1)×(2R in +1), R in is the inner diameter of the filter, indicating the radius of the target area only. The filter presents a side window shape of a hollow structure.

4. The infrared small target detection method based on edge window guided filtering according to claim 1 is characterized in that: Threshold segmentation described in step 1.3: The product of the maximum value of the target image T and the constant coefficient m is used as the segmentation threshold, and the value range of m is 0.3 to 0.5.

Images