A space-based infrared dim and small moving target detection method

Abstract

The present invention relates to a space-based infrared dim small moving target detection method in the field of space-based remote sensing data processing technology, comprising: S1. pre-processing infrared remote sensing data to increase the contrast between the target and the background to form an infrared image sequence; S3. extracting the moving area of ​​the target using the time domain features of the target; S5. identifying and extracting potential target points in the space domain based on the Tophat background modeling method; S7. jointly processing the features of the target in the time domain and the space domain to achieve target recognition. The method of the present invention effectively suppresses the false alarm rate of target detection in traditional methods and improves the detection effect of moving dim small targets.

Description

Technical Field

[0001] The present invention relates to the technical field of space-based remote sensing data processing, and mainly to a space-based infrared dim and weak small moving target detection method, in particular to a space-based infrared dim and weak small moving target detection method combining time domain features and Tophat filtering. Background Art

[0002] Space-based detection methods mainly include optical, infrared, radar, etc. Optical detection is easily interfered by weather conditions such as clouds and fog, while radar needs to actively transmit and receive electromagnetic waves. Infrared detection is a passive optical detection that is not restricted by meteorological conditions. It has the advantages of a wide detection range and a long detection distance. It is widely used in fire warning, temperature / moisture inversion, target detection and tracking, etc.

[0003] Due to the long distance of space-based detection, the spatial resolution of infrared images is very rough, the target size is small, usually only occupies 1-2 pixels in the infrared image, and the contrast between the target and the background is very low. Due to the diffraction phenomenon, the shape of the target on the image is changeable, and the shape and texture and other features are not obvious. In addition, the inconsistent detector response causes many random noises and vector pixels in the infrared image, presenting point-like targets in the image, and the target and noise are easily confused. Therefore, detecting small targets in infrared images has always been a very difficult task.

[0004] Existing space-based infrared detection methods for dim and small moving targets can be roughly divided into two categories: spatial domain and temporal domain detection methods. Among them, the detection method based on the spatial domain mainly uses the characteristics of the target showing a Gaussian spot on the infrared image to identify the target through image background estimation, target enhancement and other methods. Common background estimation methods include median filtering, mean filtering, Gaussian filtering, two-dimensional minimum mean square filtering (TDLMS), etc. Target enhancement methods include Tophat transformation, convolution kernel regression, etc. The disadvantage of this method is that it is difficult to suppress the influence of noise on target detection. The target and noise are identified at the same time, resulting in a high false alarm rate. The other is a detection method based on the time domain. This type of method usually combines multi-frame information and considers the continuity of the target between multiple frames and the regularity of the trajectory. Compared with the detection method in the spatial domain, the false alarm rate of this type of method is significantly reduced. This type of method mainly includes multi-frame joint method, 3D matched filter method, projection transformation, energy accumulation method, time domain profile, etc. However, among these methods, the multi-frame joint method, 3D matched filtering and other methods have large computational complexity and low computational efficiency; the projection transformation method will lose the signal-to-noise ratio (SNR) of the target signal, which brings difficulties to the target detection after projection; the energy accumulation method can obtain the trajectory information of the target, but it is difficult to obtain the trajectory information of weak targets in a single frame; the time domain profile method realizes the spectral changes caused by the target in the time domain when it passes through the current frame, but the feature extraction of the moving target is not significant enough and the extraction effect is not ideal. Summary of the invention

[0005] In order to overcome the defects in the above-mentioned prior art, the purpose of the present invention is to provide a space-based infrared dim small moving target detection method, which effectively suppresses the false alarm rate of target detection by traditional methods and improves the detection effect of moving dim small targets.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution of the present invention is:

[0007] The present invention provides a space-based infrared dim and small moving target detection method, comprising:

[0008] S1. Preprocess the infrared remote sensing data to increase the contrast between the target and the background and form an infrared image sequence;

[0009] S3. Extract the target's motion area by using the target's time domain features;

[0010] S5. Based on the Tophat background modeling method, identify and extract potential target points in the spatial domain;

[0011] S7. Jointly process the target features in the time domain and space domain to achieve target recognition.

[0012] According to one aspect of the present invention, step S1 comprises:

[0013] S11. Determine the area range of the ROI (Region of Interest) box in the infrared image, and use the ROI data to crop the target area of ​​the infrared image frame by frame;

[0014] S12. Data layer overlay, the cropped image data is overlaid frame by frame according to the spatial range to form an infrared image sequence;

[0015] S13. Read the infrared images in sequence into a three-dimensional matrix to form a data set containing spatial domain and time domain information.

[0016] According to one aspect of the present invention, step S3 comprises:

[0017] S31. Smoothing the infrared image sequence data in the time domain to remove outliers;

[0018] S32. Based on time domain smoothing, perform time domain profile processing to analyze the mean, variance and metric parameters of pixels in the time series;

[0019] S33. According to the dynamic change law between the target and background pixels, the key features of the target in the time domain are constructed and fitted;

[0020] S34. Based on the key features in the time domain, feature points are fitted to generate a fitting curve, motion parameters of the target are extracted and thresholded, and a motion area map of the target is obtained.

[0021] According to one aspect of the present invention, in step S32, the calculation formula of the mean is:

[0022]

[0023] The calculation formula of the variance is:

[0024]

[0025] The calculation formula of the metric parameter is:

[0026]

[0027] Where, T(i) represents the gray value (DN value) of the target in the i-th frame of infrared image; N represents the number of infrared image sequences; μ represents the mean value of the pixel in the time series; σ 2 represents the variance of the pixel in the time series; max() represents the maximum value function; η represents the measurement parameter of the target in the time domain.

[0028] According to one aspect of the present invention, step S33 comprises:

[0029] S331. In the time domain, using the maximum function, extracting potential target points;

[0030] S332. According to the different stages of the time series frame where the potential target point is located, extract and construct corresponding time domain features.

[0031] According to one aspect of the present invention, in the process of the potential target point entering and leaving the pixel, the DN value shows a change pattern of first changing from small to large and then changing from large to small, which is an important prerequisite for extracting potential target points based on time series features. The extracted potential target point is at the maximum value of the curve. When the maximum value, that is, the moment when the potential target point appears, is at the end of the time series frame, there will be a lack of change information when the potential target point leaves the pixel, and a symmetrical generation method is used to obtain a complete feature change curve; when the maximum value, that is, the potential target point appears at the beginning of the time series frame, due to the lack of change information when the potential target point enters the pixel, the symmetrical generation method is also used to obtain a complete feature change curve; when the potential target point is in other stages of the time series frame, the feature is directly extracted.

[0032] According to one aspect of the present invention, in step S34, the fitting curve is:

[0033] y=ax 2 +bx+c

[0034] Among them, a is the coefficient of the quadratic polynomial, which represents the size of the curve opening, represents the movement speed characteristics of the target, and is the main motion parameter that distinguishes the target from the background; b is the coefficient of the linear polynomial, which together with a determines the position of the central symmetry axis; c represents the constant term, which together with a and b determines the vertex position.

[0035] According to one aspect of the present invention, step S5 comprises:

[0036] S51. Construct the Tophat kernel structure based on the morphological features of the target to accurately identify the target;

[0037] S52. Using the constructed Tophat kernel structure, the original image is Tophat transformed to obtain a Tophat image background model;

[0038] S53. Subtract the Tophat image background model from the original image to generate a residual image, and obtain the potential target in the spatial domain by setting a threshold.

[0039] According to one aspect of the present invention, in step S52, the process of the Tophat transformation is: firstly, the original image is dilated, then eroded, and finally a Tophat image background model is obtained.

[0040] According to one aspect of the present invention, step S7 comprises:

[0041] S71. Superimposing the target's motion area and the potential target point to filter out false alarms caused by noise points;

[0042] S72. Analyze the spectral characteristics of the target, set a threshold for the brightness of the vector pixel, and filter out the noise points of the vector pixel;

[0043] S73. Remove the false target points to obtain the moving target of the infrared image;

[0044] S74. Perform motion target detection on the infrared image sequence frame by frame, associate the detection results with multiple frames, and use the Kalman filter tracking algorithm to locate and track the moving target to obtain the final extraction result.

[0045] Beneficial effects:

[0046] According to the solution of the present invention, the influence of high-reflection objects on the surface on the target is suppressed by data preprocessing, and the contrast between the target and the background is enhanced. The target motion area is extracted by analyzing the time domain characteristics of the pixel. Based on Tophat background modeling, potential target points in the spatial domain are obtained. Through joint processing in the time and space domains, the target detection effect is improved by combining the spectral, morphological and other characteristics of the target, and the motion tracking of the target between video frames is achieved through Kalman filtering. The space-based infrared dim and weak small moving target detection method of the present invention is based on the joint time domain characteristics and Tophat filtering, which effectively suppresses the false alarm rate of target detection and improves the detection effect of moving dim and weak small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flowchart schematically showing a method for detecting a space-based infrared dim and small moving target according to an embodiment of the present invention;

[0048] Figure 2 A specific flow chart schematically showing a method for detecting a space-based infrared dim and small moving target according to an embodiment of the present invention;

[0049] Figure 3 Schematically showing an image data diagram for target tracking based on GF-4 simulation of a space-based infrared dim and small moving target detection method according to an embodiment of the present invention;

[0050] Figure 4 Schematically showing a characteristic curve and a fitting curve of a target in the time domain in a space-based infrared dim and small moving target detection method according to an embodiment of the present invention;

[0051] Figure 5A schematic diagram of a target motion region based on time domain feature extraction in a space-based infrared dim and small moving target detection method according to an embodiment of the present invention is schematically shown;

[0052] Figure 6 A schematic diagram schematically shows a potential target diagram based on spatial feature extraction in a space-based infrared dim and small moving target detection method according to an embodiment of the present invention;

[0053] Figure 7 The schematic diagram shows the positioning and identification of the target in the image by combining the spatiotemporal domain features in a space-based infrared dim and small moving target detection method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0055] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments cannot be described one by one here, but the embodiments of the present invention are not therefore limited to the following embodiments.

[0056] The space-based infrared dim and small moving target detection method of the present embodiment combines time domain features and Tophat filtering, in order to solve the problem that the target features in space-based infrared images are not obvious and the targets are easily confused with noise such as vector pixels, which makes it difficult to detect and identify space-based infrared dim and small moving targets. On the basis of in-depth mining of the motion features of the moving targets in the time domain, a motion feature index of the moving targets is constructed, and the moving targets are accurately identified in combination with the time and space domain features, which effectively suppresses the false alarm rate of detection in the existing target detection methods and improves the detection effect of small moving targets.

[0057] Figure 1 , Figure 2 The general execution flow of each step of the space-based infrared dim and small moving target detection method of this embodiment and the specific execution steps and processes in each step are schematically represented respectively.

[0058] like Figure 1 and Figure 2 As shown in Figure 1, the space-based infrared dim and small moving target detection method includes four processes: data preprocessing, motion region extraction in the time domain, potential target extraction in the space domain, and joint processing in the time and space domains. Figure 1 As shown, the implementation steps of the method include the following:

[0059] S1. Preprocess the infrared remote sensing data to increase the contrast between the target and the background and form an infrared image sequence;

[0060] S3. Extract the target's motion area by using the target's time domain features;

[0061] S5. Based on the Tophat background modeling method, potential target points are identified and extracted in the spatial domain;

[0062] S7. Jointly process the target features in the time domain and space domain to achieve target recognition.

[0063] In this embodiment, the public infrared band data of the GF-4 satellite is used as the data source. Taking the ship target as an example, infrared sequence images and moving targets are generated through data simulation, and the above steps and their specific processes are explained and described.

[0064] Step 1: Preprocess the infrared remote sensing data to increase the contrast between the target and the background to form an infrared image sequence. The spectral difference of the original remote sensing image is too large. Bare land, beaches and other objects show high reflectivity, and the background shows low reflectivity, which makes the target easily submerged in the background and difficult to identify. Therefore, it is necessary to crop the target area first to avoid the impact of high reflectivity objects such as beaches and bare land on the target and increase the contrast between the target and the background. The specific process includes:

[0065] First, the target area is cropped. The ROI (Region of Interest) box area in the infrared image is determined, and the target area of ​​the infrared image is cropped frame by frame using the ROI data.

[0066] Figure 3 The following is a schematic diagram of the image data for target tracking based on GF-4 simulation in this embodiment. In this embodiment, a 200*200 ocean area is first cut out frame by frame from the simulated infrared image sequence as the research area. The time series range is 100 frames, and the detection of moving targets is carried out in this time period and in this area. The simulation research area results are shown in Figure 2. Figure 3 shown.

[0067] Secondly, the data layers are superimposed, and the cropped image data is superimposed frame by frame according to the time series and spatial range to form an infrared image sequence. Then, the infrared image sequence is processed and the infrared images are read into a three-dimensional matrix in sequence to form a data set containing spatial and temporal domain information.

[0068] Step 2: Comprehensively utilize the statistical characteristics of the target in the time domain to extract the target's motion area.

[0069] The specific process includes:

[0070] First, the infrared image sequence data is smoothed in the time domain to remove the influence of outliers. Secondly, based on the time domain smoothing, the time domain profile processing is performed to analyze the mean, variance and metric parameters of the pixels in the time series. Among them, the calculation formula of the mean is:

[0071]

[0072] The formula for calculating variance is:

[0073]

[0074] The calculation formula of the measurement parameter is:

[0075]

[0076] Where, T(i) represents the gray value (DN value) of the target in the i-th frame of infrared image; N represents the number of infrared image sequences; μ represents the mean value of the pixel in the time series; σ 2 represents the variance of the pixel in the time series; max() represents the maximum value function; η represents the measurement parameter of the target in the time domain.

[0077] Then, according to the dynamic change law between the target and background pixels, the key features of the target in the time domain are constructed and fitted. In the time domain characteristics of the target, when the target moves toward the background pixel, the grayscale value of the background pixel will show a dynamic change from the background grayscale value to the target grayscale value, and when the target leaves the background pixel, it will show a dynamic change from the target grayscale value to the background grayscale value. This change in the time domain is expressed as an approximate quadratic equation curve. In summary, in the process of the target entering and leaving the pixel, the DN value will show a change law from small to large, and then from large to small. This is an important prerequisite for extracting the target based on time series features. Therefore, the key features of the target are constructed. The details are as follows:

[0078] Potential target points in the time domain: In the time domain, the maximum function is used to extract potential target points. The extracted potential target points are the maximum values ​​of the above quadratic equation curve. Potential target points are local maximum points.

[0079] Time domain feature construction: As mentioned above, when the target passes through and leaves the pixel, it is expressed as a quadratic equation curve that changes from the background value to the target value and then from the target value to the background value. The process of feature construction is the process of finding this change. It should be noted that the targets or target points generated in the time domain or space domain can be understood as potential targets or potential target points. When the potential target point appears at the end of the time series frame, due to the lack of change information when the potential target point leaves the pixel, a symmetrical generation method is used to generate a complete feature change curve. Similarly, when the maximum value, that is, the potential target point, appears at the beginning of the time series frame, due to the lack of change information when the potential target point enters the pixel, a symmetrical generation method is also used to obtain a complete feature change curve. When the potential target point appears in other areas, the feature of this segment is directly extracted.

[0080] Next, based on the construction of key features in the time domain, a quadratic equation is used to fit the feature points, generate a fitting curve, extract the target's motion parameters and perform threshold processing to obtain the target's motion area map. The schematic diagram of the target's time domain feature curve and the fitting line is shown in Figure 2. Figure 4 As shown, the points constitute the characteristic curve, and the solid line is the fitting curve. The motion parameter threshold selected in this implementation is [-0.8, 0.2], and the target motion area obtained is as follows Figure 5 The fitting curve is a quadratic curve equation, as shown below:

[0081] y=ax 2 +bx+c

[0082] Among them, a is the coefficient of the quadratic polynomial, which represents the size of the curve opening, represents the movement speed characteristics of the target, and is the main motion parameter that distinguishes the target from the background; b is the coefficient of the linear polynomial, which together with a determines the position of the central symmetry axis; c represents the constant term, which together with a and b determines the vertex position.

[0083] Step 3: Based on the Tophat background modeling method, potential target points are identified and extracted in the spatial domain. Due to the influence of factors such as position change and light diffraction, the shape and texture features of the target in the infrared image are not obvious, but the overall characteristics are similar to those of a Gaussian spot. Therefore, potential target points are identified through the Tophat background modeling method. The specific steps are as follows:

[0084] First, a Tophat kernel structure is constructed for the morphological features of the target to accurately identify the target. Tophat transformation is a morphological operation, and the kernel of the morphological operation is of great significance to the morphological operation. Therefore, it is necessary to construct a kernel structure for the morphological features of the target to identify the target more accurately. In this implementation, according to the characteristics of the target, the kernel structure set is a ring structure with parameters of 9 and 4.

[0085] Secondly, using the constructed Tophat kernel structure, the original image is transformed by Tophat. The original image is first dilated and then eroded to obtain the Tophat image background model.

[0086] Next, based on the image background model constructed by Tophat, the original image is subtracted from the Tophat image background model to generate a residual image, and the potential target in the spatial domain is obtained by setting an appropriate threshold. Since the residual image contains the target, it also contains vector pixels and random noise points. Therefore, the potential target points obtained based on the spatial domain features contain a large number of false alarms. The spatial domain target extraction results are shown in Figure 2. Figure 6 shown.

[0087] Step 4: Combine the time and space domains of the target to achieve target recognition. Combine the time and space domains of the moving target to identify the target and display the final result. The specific steps are as follows:

[0088] First, the target motion area obtained in the second step and the potential target points obtained in the third step are superimposed to effectively filter out false alarms caused by noise points.

[0089] Secondly, analyze the spectral characteristics of the target, set a threshold for the brightness of the sagittal pixel, and filter out the noise points of the sagittal pixel. Although the target and noise are brighter overall in terms of grayscale characteristics, there is still a difference between the target point and the sagittal pixel. The average brightness of the sagittal pixel is higher than that of the target point. Therefore, according to this feature, a suitable threshold can be set to filter out noise points such as sagittal pixels.

[0090] Then, the spatial neighborhood features are analyzed. Since the shape and texture features of the target are not obvious and changeable, they are usually shown as several connected pixels on the infrared image. Therefore, based on this characteristic, the false target points with too many connected pixels can be removed to obtain the moving target on the infrared image.

[0091] Next, the infrared image sequence is subjected to motion target detection frame by frame, the detection results are correlated across multiple frames, and the Kalman filter tracking algorithm is used to locate and track the moving target to obtain the final extraction result. The final result obtained at this time is the final moving target detected by this method. The positioning and tracking effect of the moving target in the original infrared image is shown in Figure 2. Figure 7 shown.

[0092] The above is only one embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A space-based infrared dim and small moving target detection method, comprising: S1. Preprocess the infrared remote sensing data to increase the contrast between the target and the background to form an infrared image sequence; S3. Using the time domain features of the target, extracting the motion area of ​​the target; S5. Based on the Tophat background modeling method, potential target points are identified and extracted in the spatial domain; S7. Jointly processing the characteristics of the target in the time domain and the space domain to achieve identification of the target; The step S3 comprises: S31. Smoothing the data of the infrared image sequence in the time domain to remove outliers; S32. Based on time domain smoothing, time domain profile processing is performed to analyze the mean, variance and metric parameters of the pixels in the time series; the calculation formula of the mean is: The calculation formula of the variance is: The calculation formula of the metric parameter is: Where T(i) represents the gray value of the target in the i-th frame of infrared image; N represents the number of infrared image sequences; μ represents the mean value of the pixel in the time series; σ 2 represents the variance of the pixel in the time series; max() represents the maximum value function; η represents the measurement parameter of the target in the time domain; S33. According to the dynamic change law between the target and background pixels, construct and fit the key features of the target in the time domain; S34. Generate a fitting curve based on the key feature fitting feature points, extract the motion parameters of the target and perform threshold processing to obtain a motion area map of the target.

2. The target detection method according to claim 1, characterized in that: The step S1 comprises: S11. Determine the area range of the ROI frame in the infrared image, and use the ROI data to crop the target area of ​​the infrared image frame by frame; S12. The cropped image data are superimposed frame by frame according to the spatial range to form an infrared image sequence; S13. Read the infrared images in sequence into a three-dimensional matrix to form a data set containing spatial domain and time domain information.

3. The target detection method according to claim 1, characterized in that: The step S33 comprises: S331. In the time domain, using the maximum function, extracting potential target points; S332. According to the different stages of the time series frame where the potential target point is located, extract and construct corresponding time domain features.

4. The target detection method according to claim 3, characterized in that: The step S332 includes: when the potential target point appears at the end of the time series frame, a symmetrical generation method is used to obtain a complete feature change curve; when the potential target point appears at the beginning of the time series frame, a symmetrical generation method is used to obtain a complete feature change curve; when the potential target point is in other stages of the time series frame, features are directly extracted.

5. The target detection method according to claim 1, characterized in that: In step S34, the fitting curve is: y=ax 2 +bx+c Among them, a is the coefficient of the quadratic polynomial, which represents the motion speed characteristics of the target and is the main motion parameter that distinguishes the target from the background; b is the coefficient of the linear polynomial; c represents the constant term.

6. The target detection method according to claim 1, characterized in that: The step S5 comprises: S51. Constructing a Tophat kernel structure based on the morphological features of the target to accurately identify the target; S52. Using the constructed Tophat kernel structure, the original image is Tophat transformed to obtain a Tophat image background model; S53. Subtract the Tophat image background model from the original image to generate a residual image, and obtain potential targets in the spatial domain by setting a threshold.

7. The target detection method according to claim 6, characterized in that: The Tophat transformation in step S52 includes: firstly performing an expansion operation on the original image, and then performing an erosion operation, so as to obtain the Tophat image background model.

8. The target detection method according to claim 1, characterized in that: The step S7 comprises: S71. Superimposing the target's motion area and the potential target point to filter out false alarms caused by noise points; S72. Analyze the spectral characteristics of the target, set a threshold for the brightness of the vector pixel, and filter out the noise points of the vector pixel; S73. Remove false target points through spatial neighborhood feature analysis to obtain moving targets in infrared images; S74. Perform motion target detection on the infrared image sequence frame by frame, associate the detection results with multiple frames, and use a Kalman filter tracking algorithm to locate and track the moving target to obtain a final extraction result.

Images