A real-time infrared weak and small target detection method, device and computer equipment

Abstract

The application discloses a kind of real-time infrared weak small target detection method, device and computer equipment, belong to infrared small target detection technical field.Its method includes: according to the ratio and difference value operation of image pixel, combine RLCM and DLCM, generate RDLCM;According to the average gray of maximum pixel and the average of all pixels, generate IRIL;And according to the ratio and difference value operation of image pixel of IRIL, generate WLCM;RDLCM is weighted by WLCM, generate WRDLCM;Using multi-scale algorithm, the original image is input into WRDLCM, and maximum pooling is carried out between different scales, to obtain the maximum value of WRDLCM under three scales;The pixel corresponding to the maximum value of WRDLCM is compared with threshold value, and the pixel greater than threshold value is regarded as target pixel, and the connected region of target pixel is infrared weak small target in original image.The application combines the advantages of ratio form method and difference form method, can suppress all types of interference while enhancing different size of real target, and does not need any pre-algorithm.

Description

Technical Field

[0001] This invention relates to the field of infrared small target detection technology, and in particular to a real-time infrared weak target detection method, apparatus and computer equipment based on weighted ratio difference combined with local contrast. Background Technology

[0002] Infrared (IR) small target detection plays a crucial role in precision guidance, early warning, and maritime target search, and is even more important within a single frame. However, achieving real-time infrared small, dim target detection with high detection rate and low false alarm rate within a single frame is often a challenging task due to the following facts: a) In the acquired infrared image, the target is often far from the infrared detector, occupies only a few pixels, and is dark in grayscale, making it impossible to utilize any shape or texture information, resulting in a low detection rate; b) Some background areas may have higher brightness than the real target, which may obscure the real target and lead to a high false alarm rate; c) Complex background edges and pixel-sized noise (PNHB) with high brightness may be falsely detected as targets, also resulting in a high false alarm rate.

[0003] Algorithms for detecting small, dim infrared targets already exist, including spatial domain algorithms, frequency domain algorithms, morphological algorithms, and background estimation algorithms. Learning-based algorithms, including supervised and unsupervised types that rely on prior knowledge, have also been carefully studied. In recent years, the local contrast mechanism of the human visual system (HVS) has been introduced into the field of infrared small, dim target detection and has achieved good detection performance. Existing local contrast-based algorithms can be divided into three categories: spatial local contrast, temporal local contrast, and spatiotemporal combined local contrast. The latter two categories (such as temporal contrast filters (TCF) and spatiotemporal local contrast methods (STLCM)) require the calculation of temporal local contrast information across multiple frames, so they cannot output the target in real time. Spatial local contrast algorithms tend to calculate local contrast information within a single frame, thus they perform better in terms of detection speed.

[0004] Currently, numerous spatial local contrast algorithms exist, which can accurately locate targets by calculating the spatial local contrast between the target region and its adjacent background. Examples include the Laplacian Gaussian (LoG) filter, the Difference of Gaussians (DoG) filter, Local Contrast Measurement (LCM), Improved LCM (ILCM), Novel LCM (NLCM), Relative LCM (RLCM), Multi-scale Patch Contrast Measurement (MPCM), Multi-directional Two-dimensional Least Mean Square (MDTDLMS), and Multi-scale Three-layer LCM (TLLCM). These can be categorized into ratio-based and difference-based algorithms. Theoretically, ratio-based algorithms (such as LCM and ILCM) can effectively enhance realistic targets but cannot effectively eliminate high-brightness backgrounds; difference-based algorithms (such as LoG and DoG) can effectively eliminate high-brightness backgrounds but cannot effectively enhance realistic targets.

[0005] In recent research, local contrast algorithms with weighted functions are becoming increasingly popular. Appropriate weighting functions can significantly improve detection performance. Some researchers directly choose local statistics as weighting functions. For example, they use local entropy as a weighting function, use the variance of the central cell, utilize the standard deviation of the central cell, and employ local SCR, etc. Other researchers prefer to define and compute their own weighting functions. For example, they use the difference in local region similarity as a weighting function, local self-similarity, derivative entropy, the difference between the central and surrounding variances, consider the number of surrounding bright pixels, and use the region intensity level (RIL) to assess the complexity of a cell, then use the difference in RIL between the center and the surrounding area as a weighting function, etc.

[0006] Generally, local contrast algorithms with weighted functions typically achieve better detection performance than basic local contrast algorithms because they consider more information. However, existing weighted local contrast algorithms still have some drawbacks. First, they choose ratio-based or difference-based algorithms as the basic local contrast algorithms. Second, when calculating the weighting function, they only use difference operations and do not consider ratio operations. Third, some weighting functions are sensitive to random noise; for example, RIL uses the largest pixel value in a cell, but if the largest pixel is random noise, the result will be confused with the result of the true target. Summary of the Invention

[0007] Therefore, it is necessary to provide a real-time infrared weak target detection method, device, and computer equipment to address the above-mentioned technical problems.

[0008] This invention provides a real-time infrared weak target detection method, comprising:

[0009] Acquire raw images of weak infrared targets;

[0010] Based on the ratio and difference calculations of image pixels, and combining the ratio-based spatial local contrast algorithm RLCM and the difference-based spatial local contrast algorithm DLCM, a ratio-difference joint spatial local contrast algorithm RDLCM is generated.

[0011] An improved region intensity level (IRIL) is generated based on the average gray level of the largest pixel and the average value of all pixels; and an improved weighting function (WLCM) is generated based on the image pixel ratio and difference calculated based on the improved region intensity level (IRIL).

[0012] The weighted comparison difference joint spatial local contrast algorithm RDLCM is generated by weighting the improved weighted function WLCM with the comparison difference joint spatial local contrast algorithm RDLCM.

[0013] A multi-scale algorithm is adopted. The original image is input into the weighted ratio difference joint spatial local contrast algorithm WRDLCM, and max pooling is performed between different scales to obtain the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM at three scales.

[0014] By comparing the pixels corresponding to the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM with the threshold, pixels greater than the threshold are taken as target pixels, and the connected regions of the target pixels are the infrared weak targets in the original image.

[0015] Furthermore, the original image includes: a center cell for capturing the real target; surrounding cells for capturing the surrounding background; and the cell size N is close to or larger than the size of the real target, with N set to 9×9.

[0016] Further, the expression determination step of the RDLCM algorithm includes:

[0017] The expressions for the center pixel cell(0) in the i-th direction for RLCM and DLCM are defined as follows:

[0018]

[0019] DLCM i =Imean0-Imean i , i = 1, 2, ..., 8 (2)

[0020] Where Imean0 represents the average gray level of the largest pixel in the center pixel cell(0), where the maximum gray level value K1 is the largest. i The average gray level of the surrounding pixels cell(i) with the largest gray level value K2 is represented by the following expression:

[0021]

[0022]

[0023] in, These are the j-th maximum gray values ​​of cell(0) and cell(i), respectively; and K2 is greater than K1.

[0024] The expressions for RLCM and DLCM are finally defined as follows:

[0025] RLCM = min(RLCM) i ), i = 1, 2, ..., 8 (5)

[0026] DLCM = max(0, min(DLCM) i )), i = 1, 2, ..., 8 (6)

[0027] The expression for the RDLCM of the original image is calculated by multiplying the pixel matrices of the RLCM and DLCM:

[0028] RDLCM = RLCM·DLCM. (7)

[0029] Furthermore, for the i-th direction, the expression for the improved regional intensity level IRIL is:

[0030] IRIL i =M i -mean i ,i=0,1,2,...,8 (8)

[0031] Among them, M i The mean represents the average of the K largest pixel values ​​in the surrounding pixel cell(i). i It is the average value of all pixels in the surrounding pixel cell(i).

[0032] Further, the step of determining the expression of the improved weighted function WLCM includes:

[0033] Calculate IRIL i And through max-pooling, the final IRIL value is obtained:

[0034] IRIL = max(IRIL) i ), i = 1, 2, ..., 8 (9)

[0035] By applying ratio and difference operations to the weighting function, we obtain the expression for the improved weighted function WLCM:

[0036]

[0037] WLCM = max(0, W(x, y)) (11)

[0038] Among them, the value of the improved weighting function WLCM is positive; W(x,y) is the weighting function.

[0039] Furthermore, the expression for the weighted ratio difference joint spatial local contrast algorithm WRDLCM is:

[0040] WRDLCM = RDLCM·WLCM. (12)

[0041] Furthermore, obtaining the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM at three scales includes:

[0042] A multi-scale algorithm is used for detection, setting the size of the retained layer to different scales. For each scale, the single-scale WRDLCM is calculated, and max pooling is performed between different scales.

[0043] WRDLCM = max(0, WRDLCM) p ), p = 1, 2, ..., L (13)

[0044] Where p represents the p-th scale, and L is the total number of scales.

[0045] Further, the threshold is defined as:

[0046] Th=μ+k th ×σ (14)

[0047] Where μ and σ are the mean and standard deviation of WRDLCM, respectively; k th It is a given parameter, and k th The value ranges from 2 to 7.

[0048] A real-time infrared weak target detection device, comprising:

[0049] The raw image acquisition module is used to acquire raw images of infrared targets with low density.

[0050] The RDLCM determination module is used to calculate the ratio and difference of image pixels, and combine the ratio-type spatial local contrast algorithm RLCM and the difference-type spatial local contrast algorithm DLCM to generate the ratio-difference joint spatial local contrast algorithm RDLCM.

[0051] The WLCM determination module is used to generate an improved region intensity level (IRIL) based on the average gray level of the largest pixel and the average value of all pixels; and to generate an improved weighting function (WLCM) based on the image pixel ratio and difference calculation based on the improved region intensity level (IRIL).

[0052] The WRDLCM determination module is used to generate a weighted comparison difference joint spatial local contrast algorithm WRDLCM by weighting the comparison difference joint spatial local contrast algorithm through the improved weighting function WLCM.

[0053] The multi-scale processing module is used to input the original image into the weighted ratio difference joint spatial local contrast algorithm WRDLCM using a multi-scale algorithm, and to perform max pooling between different scales to obtain the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM at the three scales.

[0054] The target detection module compares the pixel corresponding to the maximum value of the weighted ratio-difference joint spatial local contrast algorithm WRDLCM with a threshold, and takes the pixel greater than the threshold as the target pixel, and the connected region of the target pixel is the infrared weak target in the original image.

[0055] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0056] Acquire raw images of weak infrared targets;

[0057] Based on the ratio and difference calculations of image pixels, and combining the ratio-based spatial local contrast algorithm RLCM and the difference-based spatial local contrast algorithm DLCM, a ratio-difference joint spatial local contrast algorithm RDLCM is generated.

[0058] An improved region intensity level (IRIL) is generated based on the average gray level of the largest pixel and the average value of all pixels; and an improved weighting function (WLCM) is generated based on the image pixel ratio and difference calculated based on the improved region intensity level (IRIL).

[0059] The weighted comparison difference joint spatial local contrast algorithm RDLCM is generated by weighting the improved weighted function WLCM with the comparison difference joint spatial local contrast algorithm RDLCM.

[0060] A multi-scale algorithm is adopted. The original image is input into the weighted ratio difference joint spatial local contrast algorithm WRDLCM, and max pooling is performed between different scales to obtain the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM at three scales.

[0061] By comparing the pixels corresponding to the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM with the threshold, pixels greater than the threshold are taken as target pixels, and the connected regions of the target pixels are the infrared weak targets in the original image.

[0062] The real-time infrared weak target detection method, apparatus, and computer equipment provided in the embodiments of the present invention have the following advantages compared with the prior art:

[0063] This invention proposes a weighted ratio-difference joint local contrast method (WRDLCM). It can be divided into two modules: a basic local contrast algorithm and a weighting function. The basic local contrast algorithm employs a ratio-difference combined local contrast algorithm to enhance the real target while suppressing high-brightness backgrounds. The weighting function introduces the idea of ​​ratio-difference combination, using ratio and difference operations to calculate the weighting function, thereby enhancing the real target and further suppressing high-brightness backgrounds. Specifically, an improved RIL (IRIL) is proposed before calculating the weighting function, which uses the average of some maximum pixels instead of a single maximum pixel, thus suppressing random noise. Experiments on some real infrared sequences show that the proposed WRDLCM achieves better detection performance in complex backgrounds. Attached Figure Description

[0064] Figure 1 A flowchart of a real-time infrared weak target detection method provided in one embodiment;

[0065] Figures 2a-1 to 2a-6 The original image samples corresponding to sequences 1 to 6 provided in one embodiment;

[0066] Figures 2b-1 to 2b-6 Here are the RLCM results for scale 1 provided in one embodiment;

[0067] Figures 2c-1 to 2c-6 The DLCM results for scale 1 are provided in one embodiment;

[0068] Figures 2d-1 to 2d-6 The result of RDLCM at scale 1 is provided in one embodiment;

[0069] Figures 2e-1 to 2e-6 The WLCM results at scale 1 are provided in one embodiment;

[0070] Figures 2f-1 to 2f-6 The results of WRDLCM at scale 1 are provided in one embodiment;

[0071] Figure 2g-1~2g-6 To suppress complex background results in the WRDLCM results provided in one embodiment;

[0072] Figures 2h-1 to 2h-6 The true target of the WRDLCM results provided in one embodiment;

[0073] Figures 3a-3f This is a ROC curve diagram corresponding to sequences 1 to 6 provided in one embodiment. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0075] In one embodiment, a real-time infrared weak target detection method is provided, which specifically includes:

[0076] First, the RDLCM contrast algorithm based on a combination of ratio and difference is analyzed. Second, a novel weighted function WLCM is proposed to suppress complex background and point noise. Finally, a multi-scale algorithm is combined to obtain the maximum value of WLCCM at three scales, and the final detection result is obtained after calculating the threshold. The detection capability of the proposed algorithm is also analyzed.

[0077] (1) RDLCM calculation

[0078] Combining the advantages of ratio-based local contrast and difference-based local contrast, this invention proposes a ratio-difference joint local contrast algorithm (RDLCM) and provides a method for calculating the RDLCM of the original infrared image. The center pixel of cell(0) is defined for the RLCM and DLCM in the i-th direction as follows:

[0079]

[0080] DLCM i =Imean0-Imean i , i = 1, 2, ..., 8 (2)

[0081] Where Imean0 represents the average gray level of the largest pixel K1 in cell(0), Imean i The average gray level of the largest pixel in cell(i) K2 is as follows:

[0082]

[0083]

[0084] Where K1 and K2 are the maximum gray values. or It is the j-th largest gray value of cell(0) or cell(i). To obtain a larger RLCM, it is recommended to set K2 to a value slightly larger than K1.

[0085] To suppress complex background edges, directional information is utilized, and RLCM is ultimately defined as follows:

[0086] RLCM = min(RLCM) i ), i = 1, 2, ..., 8 (5)

[0087] Similar to RLCM, DLCM is ultimately defined as

[0088] DLCM = max(0, min(DLCM) i )), i = 1, 2, ..., 8 (6)

[0089] The RDLCM of the original infrared image is calculated by multiplying the pixel matrices of the RLCM and DLCM.

[0090] TDLCM = RLCM·DLCM (7)

[0091] It's important to note that WRDLCM describes pixel location by generating signal values. A small target is salient in a local region, not across the entire image, so a small local image patch is considered, utilizing the ratio and difference between cell(0) and cell(i) (i = 1, 2, ..., 8 representing different directions). Here, the center cell is used to capture the real target, and the surrounding cells are used to capture the surrounding background; therefore, the cell size N should be close to or slightly larger than the real target. If the target size is unknown, to ensure a small cell can contain the entire target while introducing as little interference as possible, N should approximate the typical maximum size of a small target. According to the Society of Optical Instrument Engineers (SPIE), small targets are typically smaller than 9×9, so it is recommended that N be set to around 9×9. The current pixel (x, y) is placed at the center of the center cell.

[0092] (2) Calculation of WLCM

[0093] This invention proposes a new weighting function, which consists of two parts: one part is IRIL, and the other part is the calculation of the ratio-difference joint weighting function.

[0094] 1) Improved RIL

[0095] RIL is an efficient method for evaluating sub-block complexity; however, the original RIL is simply defined as the difference between the maximum and average values ​​of a sub-block and is sensitive to individual random noise. In this invention, the proposed IRIL is defined as:

[0096] IRIL i =M i -mean i ,i=0,1,2,...,8 (8)

[0097] Among them, M i The mean represents the average gray level of the largest pixel value k in cell(i).i It is the average value of all pixels in cell(i).

[0098] 2) WLCM

[0099] IRIL = max(IRIL) i ), i = 1, 2, ..., 8 (9)

[0100] Calculate IRIL i Then, the final IRIL value is obtained through max-pooling.

[0101]

[0102] In this invention, both ratio and difference operations are used in the weighting function:

[0103] WLCM = max(0, W(x, y)) (11)

[0104] When defining the WLCM, non-negativity constraints are considered and the weight function is ensured to be positive. Compared with some existing weight function definitions, the weight function used in this invention utilizes the characteristics of the target and the background, as well as the differences between the two, thus providing a more comprehensive consideration.

[0105] (3)WRDLCM

[0106] This paper proposes a new algorithm called WRDLCM, which combines RDLCM and the weight function WLCM. Its definition is...

[0107] WRDLCM=RDLCM·WLCM (12)

[0108] (4) Multiscale detection

[0109] K1 and K2 in equations (3) and (4) are key parameters in the proposed algorithm. To obtain better detection performance, K1 and K2 should be adaptively adjusted according to the size of the target. However, in practical applications, the target size is usually unknown. Therefore, the algorithm proposed in this paper adopts multi-scale detection, sets the size of the retention layer to different scales, and calculates the single-scale WRDLCM according to equations (1) to (12) for each scale, and then performs max pooling between different scales:

[0110] WRDLCM = max(0, WRDLCM) p ), p = 1, 2, ..., L (13)

[0111] Where p represents the p-th scale, and L is the total number of scales. This invention assigns different K1, K2, and N values ​​based on different P values, where N is the size of the sub-block. Note that nonnegative constraints are used to further suppress clutter, as the pixel values ​​of the true target are typically brighter than their neighbors.

[0112] (5) Threshold operation

[0113] For each pixel of the original image, multi-scale WRDLCM is computed and the results are formed into a new matrix called a saliency map (SM). It is necessary to discuss the different cases when pixel (x,y) is the center of the true object, pure background, background edge, broken cloud, or PNHB.

[0114] a) If (x,y) is the true target center, then its Imean0 and WLCM will be very large. i It will be very small, so its WRDLCM will be greater than 0.

[0115] b) If (x,y) is a pure background (including a bright background), since the background is usually continuously distributed over a large area, its Imean0 will be close to Imeani. If (x,y) is a background edge, then WRDLCM≈0.

[0116] c) If the minute operation is used in equation (6), and if (x,y) is a fragmented cloud, then TLLCM≈0.

[0117] d) Because fragmented clouds are usually locally prominent, their Imean0 may be greater than Imean. i Therefore, its RDLCM value will be greater than 0. However, the grayscale value of a broken cloud is usually smaller than that of the real target, so its RDLCM will be smaller than that of the real target, and it will not interfere with target detection.

[0118] (e) If (x,y) is a PNHB with a grayscale value similar to that of the real target, since PNHB is usually displayed as a single pixel, its Imean0 and WLCM will be smaller than those of the real target, and it will not interfere with the detection.

[0119] There are many definitions of threshold. This invention adopts the widely used idea of ​​Gaussian thresholding because it is an adaptive threshold definition that can effectively capture anomalous and significant information in large amounts of data, and adaptively defines the threshold as follows:

[0120] Th=μ+k th ×σ (14)

[0121] Where μ and σ are the mean and standard deviation of WRDLCM, and k thIt is a given parameter. Experiments show that k th The optimal range is from 2 to 7. In WRDLCM, pixels with a Th value greater than Th will be output as target pixels, while other pixels will be discarded. In the final detection result, each connected region is considered as a detected target (dilation may be necessary to eliminate clutter).

[0122] Experimental Results and Analysis

[0123] The flowchart of the entire detection algorithm is as follows: Figure 1 As shown. For each scale, the rlcm, dlcm, and wrld of the original infrared image are first calculated. Next, the largest wdlcm result among the three scales is output. Finally, a thresholding operation is used to extract the target. Furthermore, Figure 1 It can be seen that the algorithm has the potential for parallel processing.

[0124] To verify the effectiveness of the algorithm, a large amount of real-world image data (including sequence data and single-frame data) containing weak targets was used for experimental verification. All experimental code in this paper was run on a personal computer with a processor base frequency of 2.70 GHz, an Intel i5-6400, and 8 GB of RAM. The testing software used was MATLAB R2018b.

[0125] Table 1. Detailed information on the six infrared sequences.

[0126]

[0127] In this invention, three scales were used, with sub-block sizes set to 9×9, 7×7, and 5×5, respectively. The processing results at each stage are shown in Figure 2. Figures 2a-1 to 2a-6 These are the original image samples for each sequence. It can be seen that the target is usually very dark and small, while the background is usually very complex, with obvious edges and a lot of noise. Figures 2b-1 to 2b-6 This is the RLCM result at scale 1 (the single-scale result with a cell size of 5×5 is shown here). Figures 2c-1 to 2c-6 These are DLCM results at scale 1. Figures 2d-1 to 2d-6 The results are for scale 1 RDLCM. It can be seen that the target is enhanced and the background is suppressed in RDLCM, but there may still be some residuals with complex backgrounds. Figures 2e-1 to 2e-6 This is the WLCM result at scale 1. Figures 2f-1 to 2f-6 This is the result of WRDLCM at scale 1. It can be seen that the weight coefficients for the true target are larger, while the weight coefficients for the complex background are smaller. Therefore, as... Figure 2g-1~2g-6 As shown, the WRDLCM results can further suppress complex backgrounds and successfully detect real targets without any false alarms (Figures 2h-1 to 2h-6).

[0128] To further illustrate the effectiveness of the proposed algorithm, eight state-of-the-art algorithms were selected for comparison, including DoG, VAR-DIFF, ILCM, NLCM, MPCM, RLCM, WLDM, and MDTDLMS. DoG is a filter-based local contrast algorithm without a weighting function; VAR-DIFF is a weighted local contrast algorithm that uses the difference between the target and its neighboring background as the weighting function; ILCM and MPCM are patch-based local contrast algorithms without weighting functions; RLCM and MDTDLMS are local contrast algorithms that use both ratio and difference operations; WLDM and NLCM are weighted local contrast algorithms that use target features as the weighting function; and WRDLCM is the algorithm proposed in this paper.

[0129] In addition, Tables 2 and 3 present some features of the first frame of the six sequences, where C wh C nb The SCR is defined as follows, where is the maximum grayscale value of the target, and I... wh For the target, the maximum gray value of the entire image, I nb σ is the average gray value within the target neighborhood. wh Let S be the standard deviation of the entire image. For each pixel of the original image, calculate the WSLCM and form the results into a new matrix called the saliency map (SM), SCR. in and SCR out These are the SCRs of the original image and the SM, respectively, and σ in and σ out These are the standard deviations of the original image and the SM:

[0130]

[0131]

[0132]

[0133]

[0134] Table 2 SCRG values ​​for different algorithms

[0135]

[0136] Table 3 BSF values ​​for different algorithms

[0137]

[0138] As can be seen from Tables 2 and 3, the proposed algorithm can achieve the optimal SCRG for all six targets. Furthermore, the proposed algorithm can achieve the optimal BSF for all six sequences. This is because the proposed algorithm can purposefully enhance the shape of the real target. In addition, the ratio-difference joint weighting function in Equation (10) and the non-negativity constraint in Equation (11) can further improve the detection performance.

[0139] Then, Figures 3a-3f The ROC characteristic curves of different algorithms are presented. The false alarm rate (TPR) is defined as the ratio of the number of detected true targets to the total number of true targets, and the detection rate (FPR) is defined as the ratio of the number of detected false targets to the total number of pixels. As can be seen from the figure, the proposed algorithm can achieve the best detection performance in all cases.

[0140] In summary, single-frame infrared (IR) weak target detection, characterized by high detection rate, low false alarm rate, and high detection speed, is a challenging task because targets are typically small and dim, and subject to various types of interference, such as high-brightness backgrounds, complex background edges, and pixel-sized high-brightness noise (PNHB). Existing algorithms, such as ratio-based local contrast and difference-based local contrast algorithms, cannot effectively enhance the real target while suppressing all interferences, wasting local diversity information that could be used to further suppress complex backgrounds. Therefore, this invention proposes a weighted ratio-difference joint local contrast algorithm (WRDLCM), comprising two modules: rdlcm and wlcm. This algorithm combines the advantages of ratio-based and difference-based methods, enhancing real targets of different sizes while suppressing all types of interference, and requires no pre-algorithms. WLCM utilizes local diversity information to further suppress complex backgrounds. Experiments using six real IR sequences containing different types of backgrounds and targets of different sizes demonstrate that the proposed WRDLCM algorithm is more effective and robust than existing algorithms in terms of detection rate and false alarm rate. Furthermore, a weighting function based on the improved region intensity level (IRIL) is proposed to further suppress complex backgrounds, resulting in better suppression of random noise. Additionally, the proposed algorithm has the potential for parallel processing, which is highly valuable for improving detection speed.

[0141] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A real-time infrared weak target detection method, characterized in that, include: Acquire raw images of weak infrared targets; Based on the ratio and difference calculations of image pixels, and combining the ratio-based spatial local contrast algorithm RLCM and the difference-based spatial local contrast algorithm DLCM, a ratio-difference joint spatial local contrast algorithm RDLCM is generated. An improved region intensity level (IRIL) is generated based on the average gray level of the largest pixel and the average value of all pixels. An improved weighting function WLCM is generated based on the image pixel ratio and difference calculated using the improved regional intensity level IRIL. The weighted comparison difference joint spatial local contrast algorithm RDLCM is generated by weighting the improved weighted function WLCM with the comparison difference joint spatial local contrast algorithm RDLCM. A multi-scale algorithm is adopted. The original image is input into the weighted ratio difference joint spatial local contrast algorithm WRDLCM, and max pooling is performed between different scales to obtain the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM at three scales. By comparing the pixel corresponding to the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM with the threshold, the pixel greater than the threshold is taken as the target pixel, and the connected region of the target pixel is the infrared weak target in the original image. The original image includes: a center cell for capturing the real target; surrounding cells for capturing the surrounding background; and cell size... N If the size is close to or larger than the actual target size, N is set to 9×9; The expression determination steps for the RDLCM (Relational Local Contrast Comparison) algorithm include: center pixel cell ( 0 For the first i The expressions for directional RLCM and DLCM are defined as follows: in, I mean 0 Represents the center pixel cell(0) Maximum gray value in the center K 1 The maximum average grayscale of pixels. I mean i Represents surrounding pixels cell(i) Maximum gray value around the center K 2 The maximum average gray level of pixels is expressed as follows: in, , They are cell(0) , cell(i) The j The maximum grayscale value; and K 2 Greater than K 1 ; The expressions for RLCM and DLCM are finally defined as follows: The expression for the RDLCM of the original image is calculated by multiplying the pixel matrices of the RLCM and DLCM: (7) For the i The expression for the improved regional intensity level IRIL is as follows: (8) in, M i Represents surrounding pixels cell(i) Center front K The average of the maximum pixel values, mean i It is the surrounding pixels cell(i) The average value of all pixels in the array; The steps for determining the expression of the improved weighted function WLCM include: calculate IRIL i And through max-pooling, the final IRIL value is obtained: (9) By applying ratio and difference operations to the weighting function, we obtain the expression for the improved weighted function WLCM: (10) (11) Among them, the value of the improved weighted function WLCM is positive; W(x,y) It is a weighted function; The threshold is defined as: in, μ and σ These are the mean and standard deviation of WRDLCM, respectively. k th It is a given parameter, and k th The value ranges from 2 to 7.

2. The real-time infrared weak target detection method as described in claim 1, characterized in that, The expression for the weighted ratio difference joint spatial local contrast algorithm WRDLCM is: (12)。 3. The real-time infrared weak target detection method as described in claim 2, characterized in that, The process of obtaining the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM at three scales includes: A multi-scale algorithm is used for detection, setting the size of the retained layer to different scales. For each scale, the single-scale WRDLCM is calculated, and max pooling is performed between different scales. (13) in, p Indicates the first p One scale, L It is the total number of scales.

4. A real-time infrared weak target detection device based on the real-time infrared weak target detection method according to any one of claims 1 to 3, characterized in that, include: The raw image acquisition module is used to acquire raw images of infrared targets with low density. The RDLCM determination module is used to calculate the ratio and difference of image pixels, and combine the ratio-type spatial local contrast algorithm RLCM and the difference-type spatial local contrast algorithm DLCM to generate the ratio-difference joint spatial local contrast algorithm RDLCM. The WLCM determination module is used to generate an improved region intensity level (IRIL) based on the average grayscale of the largest pixel and the average value of all pixels. An improved weighting function WLCM is generated based on the image pixel ratio and difference calculated using the improved regional intensity level IRIL. The WRDLCM determination module is used to generate a weighted comparison difference joint spatial local contrast algorithm WRDLCM by weighting the comparison difference joint spatial local contrast algorithm through the improved weighting function WLCM. The multi-scale processing module is used to input the original image into the weighted ratio difference joint spatial local contrast algorithm WRDLCM using a multi-scale algorithm, and to perform max pooling between different scales to obtain the maximum value of the weighted ratio difference joint spatial local contrast algorithm WRDLCM at the three scales. The target detection module compares the pixel corresponding to the maximum value of the weighted ratio-difference joint spatial local contrast algorithm WRDLCM with a threshold, and takes the pixel greater than the threshold as the target pixel, and the connected region of the target pixel is the infrared weak target in the original image.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

Images