CN105469428B - A kind of detection method of small target based on morphologic filtering and SVD - Google Patents

Abstract

The invention discloses a weak and small target detection method based on morphological filtering and SVD. Step 1: Perform background suppression and noise removal through a morphological filtering target enhancement algorithm to obtain a preprocessed image sequence; Step 2: Read in N _max images The obtained image sequence is estimated by the number of frames to obtain the number of frames N to be processed; step 3: merge the images composed of N+1 images into two-dimensional data, find its autocorrelation matrix and perform SVD on its autocorrelation matrix; step 4: Select the appropriate feature vector to reconstruct the image sequence to obtain a new feature image sequence; Step 5: Carry out threshold segmentation on the reconstructed image sequence, and separate from the background to obtain the position of the weak target in the original image; Step 6: In the sequence Correct each image separately; Step 7: After replacing N _max with N, repeat steps 2-7. The invention effectively combines the method of morphological filtering and singular value decomposition to detect the weak and small targets in the video, has short calculation time, high detection efficiency, good accuracy and robustness.

Description

A Dim Target Detection Method Based on Morphological Filtering and SVD

technical field

The method belongs to the field of video analysis, and specifically relates to a small target detection method based on morphological filtering and SVD.

Background technique

The detection of small and weak targets is self-evident in modern warfare. At present, it has become the core technology of information processing in the fields of satellite remote sensing, high-energy physics, low-altitude early warning, and precision guidance. Due to the small number of pixels and the lack of target structure information, there is very little information available for segmentation and detection algorithms. However, the intensity of the target accepted by the sensor is weak, and the interference of noise and background clutter is strong, which reduces the signal-to-noise ratio of the image. Therefore, we should make good use of the continuity and regularity of the target in the sequence image to detect the target. For a long time, how to better use the inter-frame information of weak and small targets to improve the reliability and efficiency of detection has been the focus of weak and small target detection.

For the problem of background suppression and weak target detection, some early work mainly focused on dynamic programming and state estimation techniques to increase the detectability of the target, but it may show poor performance in the case of low signal-to-noise ratio. At present, methods such as morphological filtering, genetic algorithm, neural network algorithm, and wavelet transform have been used for weak and small target detection. However, in complex backgrounds, target points are easily overwhelmed by noise, and it is difficult to achieve reliable detection and recognition of targets. In addition, it is difficult to meet good detection performance under the conditions of high data throughput and high real-time requirements.

Contents of the invention

Purpose of the invention: Aiming at the problems existing in the prior art, the present invention provides a small target detection method based on morphological filtering and singular value decomposition (hereinafter referred to as SVD) with short calculation time and accurate detection.

SUMMARY OF THE INVENTION: The present invention provides a small target detection method based on morphological filtering and SVD, including the following steps:

Step 1: Input the video sequence to be detected, perform background suppression and noise removal through the morphological filtering target enhancement algorithm, and obtain the preprocessed image sequence;

Step 2: From the preprocessed image sequence, read in an image sequence composed of N _max images, estimate the number of frames, and obtain the number N of frames to be processed;

Step 3: Read in N+1 images, including N frames of images to be processed and the next image of N frames of images, and merge the images composed of N+1 images into two-dimensional data, find its autocorrelation matrix and Perform SVD on its autocorrelation matrix;

Step 4: Select the eigenvector corresponding to the intermediate eigenvalue to reconstruct the image sequence to obtain the reconstructed image sequence;

Step 5: Carry out threshold segmentation on the reconstructed image sequence, separate from the background to obtain the position of the weak target in the original image;

Step 6: Perform inter-frame position correction and intra-frame position correction on each image in the reconstructed image sequence obtained in step 4;

Step 7: After replacing N _max with N, repeat steps 2 to 7 until the last image in the image sequence composed of N _max images is processed and the result is output.

Further, in the step 1, the morphological filtering target enhancement algorithm uses a circle as a structural element. This will produce a better filtering effect for restoring noise-contaminated images. In this way, a better filtering effect can be obtained.

Further, in the step 1, the morphological filtering target enhancement algorithm is used to first perform a closing operation on each frame of the image in the video to be detected, and then perform a closed operation on the image that can be filled between the trachoma noise or the area that will not form a degenerated rectangle. Open operation.

Further, the method for estimating the number of frames in step 2 is: read in an image sequence composed of N _max images, perform a pairwise difference operation on all the images read in to obtain an N _max -1 frame image sequence, and then perform N _max - Sum the total pixels of each image in the 1-frame image sequence, and then divide the calculated N _max -1 pixel sum by the width and height of the corresponding image to obtain N _max -1 evaluation values, and divide N _max -1 The evaluation values are averaged to the value X of the severe degree of image change in N _max frames, and finally the severe degree of image change X is substituted into the formula , the number N of image frames to be processed this time is obtained, where N _max represents the maximum number of processed frames, and N _min represents the minimum number of processed frames.

Further, the maximum number of processing frames N _max is set to 25, and the minimum number of processing frames N _min is set to 5.

Further, in the step 5, the maximum entropy method is used to perform threshold segmentation on the image.

Further, the correction method in step 6 is: find the maximum gray value in the current search window, and then iteratively search the coordinates of the maximum gray value as the center position of the next search window until the final search window remains unchanged. The position at the center of the search window is the target position.

Beneficial effects: Compared with the prior art, the present invention effectively combines the method of morphological filtering and singular value decomposition to detect weak and small targets in the video, which not only has short calculation time and high detection efficiency, but also has high accuracy and robustness. Both are better.

Description of drawings

Fig. 1 is a work flowchart of the present invention.

Figure 2(a) and Figure 2(b) are the 20th and 50th frame images in the detection video, respectively;

Figure 3(a) and Figure 3(b) are the detection results of weak and small targets in the 20th frame and 50th frame in the detection video based on the morphological filtering method;

Figure 4(a) and Figure 4(b) are the results of detecting weak and small targets in the 20th and 50th frames of the video based on SVD, respectively;

Fig. 5(a) and Fig. 5(b) are the detection results of the weak and small targets in the 20th frame and the 50th frame in the detection video, respectively, according to the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings.

As shown in Fig. 1, the present invention provides a kind of weak target detection method based on morphological filtering and SVD, comprises the following steps:

Step 1: Input the video sequence to be detected, perform background suppression and noise removal through the morphological filtering target enhancement algorithm, and obtain the preprocessed image sequence;

Considering the detection of weak and small targets, it is more or less encountered that the noise image overlaps with the non-noise image to form agglomerates or the radius of some noise particles exceeds the radius of the non-noise particles, then in this case, a circle with a suitable radius can be selected As a structural element, this will produce a better filtering effect for restoring noise-contaminated images, because: (1) the rounding effect of the circle can play the effect of low-pass filtering; (2) the use of circular filtering does not need to consider The effect of rotation.

When determining the radius of a circular structural element, an optimization method can be used to treat the image and noise as a random process. Through statistical analysis, the particle image polluted by noise can be quantitatively analyzed, and the statistical distribution parameters can be obtained to obtain the probability of occurrence. The radius of the largest noise particle and the particle not polluted by noise is selected as the radius of the structural element, and the optimization result is obtained.

In addition, there may be both pepper noise (difference noise) and trachoma noise (combination noise) in the sequence of weak and small targets, so the effect of simply using the opening or closing operation will not be good. In order to be compatible with smooth noise and preserve image edges and other meaningful features, the present invention first performs a closing operation on each frame of image in the video to be detected, and then performs an opening operation on the area that can be opened to obtain a better filtering effect, wherein the area that can be opened is There are regions of the image that can be filled in between trachoma noise without forming severely degenerated rectangles. According to the formula and definition, it is not difficult to know that the morphological filter has good translation invariance, incrementality, idempotency and duality. These properties make the on-off or off-on filter more practical and feasible, and achieve better results.

Step 2: From the preprocessed image sequence, read in an image sequence composed of N _max images, estimate the number of frames, and obtain the number N of frames to be processed;

In the weak target sequence, the number of frames that need to be processed at one time is different for different targets, so it is necessary to obtain the number of frames N that need to be processed through adaptive determination. When choosing the number of frames, the first thing to consider is the speed of image changes, that is, when the changes are fast, select fewer frames for processing, otherwise choose more frames. Therefore, when estimating the number of frames, it is necessary to determine the maximum number of processed frames N _max and the minimum number of processed frames N _min , wherein the maximum number of processed frames N _max is the number of images N _max read in from the preprocessed image sequence. If N _max is very large, the algorithm has no real-time performance. Similarly, if N _min ≤ 0, it is meaningless to perform singular value decomposition on it. Considering real-time and feasibility, the maximum processing frame number N _max is set to 25, the minimum processing frame number N _min is set to 5, and the cosine function is used to calculate the frame number estimation between N _max and N _min , and the independent variable is the image change Intensity X.

When X<0.5, the minimum number of frames is used for processing, when X>1.5, the maximum number of frames is used for processing, and when 0.5<X<1.5, the cosine function transition method is used for processing. Using the knowledge of trigonometric functions, the formulas for obtaining the degree of image change X and the number of processing frames N are as follows:

Among them, the specific evaluation algorithm is: under the premise of N _max = 25 and N _min = 5, first read in the images with the maximum number of frames, perform pairwise difference operations on the images of these frames to obtain an image sequence of N _max -1 frames, and then Sum the total pixels of each image in the N _max -1 frame image sequence, and then divide the calculated N _max -1 pixel sum by the width and height of the corresponding image to obtain N _max -1 evaluation values, and Take the average of N _max -1 evaluation values to get the value X of the severe degree of image change in N _max frames, and finally substitute X into formula (1) to get the number N of image frames that need to be processed this time.

Step 3: Read in N+1 images, including N frames of images to be processed and the next image of N frames of images, and merge the images composed of N+1 images into two-dimensional data, find its autocorrelation matrix and Perform SVD on its autocorrelation matrix;

The image sequence A composed of N+1 images is a three-dimensional data set of {x, y, z}, where x represents the number of rows of the image, y represents the number of columns of the image, and z represents the number of frames of the sequence. Suppose the matrix B _q×z is a two-dimensional expansion of A, where q=x×y, then B is a matrix of x×y rows and z columns. In general, the value of x×y is relatively large, so the data volume of the matrix B _q×z is not suitable for direct use for singular value decomposition. Let the autocorrelation matrix of B _q×z be C _q×z , then C _q×z =B′ _q×z B _q×z , where B′ _q×z is the transpose of B _q×z , so the matrix C _q×z is a square matrix whose rows and columns are equal to z. At this time, SVD is performed on the matrix C _q×z . In order to realize the detection of weak and small targets, the above-mentioned singular value decomposition is performed on each image and its previous image sequence.

Step 4: Select the eigenvector corresponding to the intermediate eigenvalue to reconstruct the image sequence to obtain the reconstructed image sequence;

According to the results of singular value decomposition, select the appropriate eigenvector to reconstruct the corresponding image sequence, and the reconstructed image sequence is a matrix of q rows and i columns, and V _z×i is the eigenvector corresponding to the intermediate eigenvalue, which can be Transform back to image matrix r represents the total number of singular values obtained in step 3, i represents the label of the singular value, different reconstructed images can be obtained by choosing different i. Among them, the background image sequence accounts for the vast majority of the entire image sequence to be detected, and the obtained image reconstructed from the eigenvectors corresponding to the main eigenvalues will reflect the background information. The noise image sequence is a group of two-dimensional random processes, and the image reconstructed by the eigenvector corresponding to the smallest eigenvalue will reflect the noise information. The target image sequence is constantly moving and changing throughout the image sequence, and the reconstructed image obtained from the intermediate eigenvalues will enhance the moving weak target.

Step 5: Perform threshold segmentation on the reconstructed image sequence obtained in step 4, and separate the position of the weak target in the original image from the background;

Among them, in step 5, the maximum entropy method is used to threshold the image, mainly using the gray distribution density function of each image in the reconstructed image sequence to define the information entropy of the image, and different entropy is proposed according to different assumptions or different viewing angles criterion, and finally the threshold is obtained by optimizing the criterion. The information entropy of an image reflects the overall appearance of the image. If the image contains a target, the amount of information (that is, entropy) is the largest where the target and the background can be separated. In a multi-grayscale image containing multiple targets, there must be a grayscale, which can be used as a threshold to obtain the best binary segmentation of the image. In this way, the separation of the target and the background can be achieved, and its approximate position can be obtained.

Step 6: Perform inter-frame position correction and intra-frame position correction on each image in the reconstructed image sequence obtained in step 4, so as to achieve the purpose of detecting weak and small targets;

In order to improve the accuracy of target position estimation, the detection result of this time needs to be used to correct the target position when processing the next frame image and its image sequence.

The specific method is to update the target detection position according to the difference between the two different detection results before and after it is within a certain threshold, otherwise it is considered that this processing is an outlier, and the detection result is discarded. The threshold here is obtained by the method of maximum entropy.

The singular value decomposition is an image sequence, so the entire data space contains not only the data of the current image, but also the data of the previous N images. The target position obtained after reconstruction will reflect the past position of the current target to a certain extent, rather than the position of the target in the current image, so it is necessary to correct the target position within the frame.

The basic principle of the modified algorithm is: find the maximum gray value in the current search window, and then use the coordinates here as the center position of the next search window to iteratively search until the final search window remains unchanged. At this time, the position of the center of the search window is for the target location.

After SVD processing, the combination of inter-frame and intra-frame joint detection can keep the detection rate of weak and small objects in the image sequence at about 90%, which meets the real-time and feasibility requirements of image processing to a large extent.

Step 7: After replacing N _max with N, repeat steps 2 to 7 until the last image in the image sequence composed of N _max images is processed and the result is output.

As shown in Figure 2(a) and Figure 2(b), in the vs2010+opencv2.4.3 operating environment, the experiment was carried out on the actually shot sequence images of infrared weak targets. The sequence images have 100 frames in total, and the image size is 200×256 pixels , the target size is about 2×2 pixels, the signal-to-noise ratio is within 2, the contrast is 7%, and the background is the sky and clouds. Select the 20th frame image (as Figure 2(a)) and the 50th frame (Figure 2(b) ) are detected based on morphological filtering, SVD and the method provided by the present invention respectively.

As shown in Figure 3(a) to Figure 5(b), the morphological filtering method has a faster processing speed, but it cannot accurately achieve the purpose of weak and small target detection, and also detects false targets; SVD can better achieve highlighting weak and small targets , and can maintain a high detection rate, but the tracking of the target position is not ideal, and the detection calculation is heavy and the speed is slow. However, the method proposed by the present invention not only has a high detection rate, but also has a fast processing speed, and is relatively good in robustness and real-time performance.

Claims (6)

1. A weak and small target detection method based on morphological filtering and SVD, characterized in that: comprise the following steps: Step 1: Input the video sequence to be detected, perform background suppression and noise removal through the morphological filtering target enhancement algorithm, and obtain the preprocessed image sequence; Step 2: From the preprocessed image sequence, read in an image sequence composed of N _max images, estimate the number of frames, and obtain the number N of frames to be processed; Step 3: Read in N+1 images, including N frames of images to be processed and the next image of N frames of images, and merge the images composed of N+1 images into two-dimensional data, find its autocorrelation matrix and Perform SVD on its autocorrelation matrix; Step 4: Select the eigenvector corresponding to the intermediate eigenvalue to reconstruct the image sequence to obtain the reconstructed image sequence; Step 5: Carry out threshold segmentation on the reconstructed image sequence, separate from the background to obtain the position of the weak target in the original image; Step 6: Perform inter-frame position correction and intra-frame position correction on each image in the reconstructed image sequence obtained in step 4; Step 7: After replacing N _max with N, repeat steps 2 to 7 until the last image in the image sequence composed of N _max images is processed and the result is output; The method for estimating the number of frames in the step 2 is: read in an image sequence composed of N _max images, perform a pairwise difference operation on all the images read in to obtain an N _max -1 frame image sequence, and then perform N _max -1 frame The total pixels of each image in the image sequence are summed, and then divided by the obtained N _max -1 pixel sum by the width and height of the corresponding image to obtain N _max -1 evaluation values, and the N _max -1 evaluation values The value is averaged to the value X of the severe degree of image change in N _max frames, and finally the severe degree of image change X is substituted into the formula , the number N of image frames to be processed this time is obtained, where N _max represents the maximum number of processed frames, and N _min represents the minimum number of processed frames. 2. The weak and small target detection method based on morphological filtering and SVD according to claim 1, characterized in that: in the step 1, the morphological filtering target enhancement algorithm uses a circle as a structural element. 3. The weak and small target detection method based on morphological filtering and SVD according to claim 1, characterized in that: in the step 1, the morphological filtering target enhancement algorithm is used to first perform a closed operation on each frame of image in the video to be detected , and then perform an opening operation on the area that can fill the image interior between the sand hole noise or will not form a degenerate rectangle. 4. The weak and small target detection method based on morphological filtering and SVD according to claim 1, characterized in that: the maximum number of processed frames _Nmax is set to 25, and the minimum number of processed frames _Nmin is set to 5. 5. The weak and small target detection method based on morphological filtering and SVD according to claim 1, characterized in that: in said step 5, the maximum entropy method is used to perform threshold segmentation on the image. 6. The weak and small target detection method based on morphological filtering and SVD according to claim 1, characterized in that: the correction method in the step 6 is: find the maximum gray value in the current search window, and then maximize the gray value The coordinates of the value are used as the center position of the next search window to iteratively search until the final search window remains unchanged. At this time, the position of the center of the search window is the target position.