Method for Extracting and Recognizing Marine Ship Targets from Optical Remote Sensing Images

Abstract

The present invention relates to a method for extracting and recognizing marine ship targets from optical remote sensing images, including: inputting visible light remote sensing images; using a visual saliency method based on covariance joint features to globally detect the sea surface in the remote sensing images to generate a single-scale saliency map; downsampling the generated single-scale saliency map to establish a multi-scale saliency map and obtain a final saliency map; calculating the detection threshold of marine ship targets on the sea surface, binarizing the saliency map to achieve rough segmentation of the saliency map; establishing a training set and a test set; obtaining feature vectors that can be used for training and testing; training a model to classify the feature vectors according to the established framework and positive and negative samples in the training set; testing candidate regions according to the established framework and positive and negative samples in the test set to complete the extraction and recognition of marine ship targets from optical remote sensing images. The present invention can efficiently search for sea surface target areas, greatly reduce the false alarm rate, and improve the detection accuracy.

Description

Technical Field

[0001] The present invention relates to a method for extracting and identifying marine ship targets from optical remote sensing images. Background Art

[0002] In recent years, marine remote sensing technology has always been one of the challenging research topics in the field of computer vision, and among them, ship detection technology has relatively promising development prospects. With the rapid development of remote sensing information science, ship detection using remote sensing technology is not only applied to military fields such as maritime reconnaissance, maritime strike analysis and assessment, but also widely used in civilian fields such as island resource investigation, marine exploration, and maritime rescue, and has important value.

[0003] With the increasing enhancement of the remote sensing data acquisition capabilities of air and space platforms and the rapid development of high-resolution satellites, more and more remote sensing data can be used for research. From the perspective of data acquisition, current ship detection can be roughly divided into three categories: ship detection in synthetic aperture radar (SAR) images, ship detection in infrared (IR) images, and ship detection in visible light remote sensing (VRS) images. Since the synthetic aperture radar (SAR) method has the ability of day and night imaging that is not affected by complex weather conditions such as light and clouds and certain penetrability, it is mostly used to monitor oil spills on the sea surface and ocean surface currents. Countries around the world have been committed to the development of new SAR payloads, continuously improving the spatial resolution, and achieving impressive performance. However, the coherent imaging mechanism of SAR images results in a large amount of speckle noise in the images, seriously interfering with edge and texture features. In addition, color information cannot be used, so it is not suitable for identifying ship targets. In addition, infrared images are used to enhance the visual effect under low light conditions, but they also have disadvantages such as low signal-to-noise ratio and insufficient structural information. The visible light image has more available features, such as color, texture, edge, direction, and frequency domain features, so it can capture more details and complex structures.

[0004] At present, the methods for extracting and detecting the target area of maritime ships include: traditional optical remote sensing image ship detection segments the image based on gray information. Such methods are only applicable to the situation of calm sea surface without cloud and fog interference, and have poor robustness; there are also methods based on template matching. These methods mainly eliminate the offshore island area according to the shape of the ship target, but it is not easy to select templates for different scenarios and different ship types; there are also methods based on traditional machine learning, mainly to separate the target and background areas, which highly rely on ideal training samples; deep learning methods can effectively classify the target and background, but have high requirements for hardware, complex training steps, and poor interpretability; methods based on sparse representation are currently not systematic, and some are applied to ship detection in infrared images; there are also methods based on visual saliency segmentation. Compared with the number of candidate region boxes generated by the sliding window method and the gray information segmentation image method in machine learning methods, they can generate fewer suspected target areas.

[0005] All in all, for the current ship target detection task in optical remote sensing images, there are at least the following deficiencies:

[0006] 1. Islands, thick clouds, sea waves and various uncertain sea conditions lead to a high false alarm rate.

[0007] 2. The limited parameters of visible light imaging sensors, sea clutter interference, and ship wake interference reduce the homogeneity of ship targets.

[0008] 3. The factors of the ship itself such as color, texture, size and type result in low gray correlation of the target;

[0009] 4. Based on the requirement of rapidity for large-scale remote sensing data, reducing the computational burden becomes a key issue;

[0010] 5. The problem of low detection efficiency caused by various target directions and unobvious features during the detection process.

[0011] Therefore, how to quickly, stably and robustly extract and detect in complex situations such as bad sea conditions, low target homogeneity and gray correlation, target geometric distortion, and target rotation has become an urgent problem to be solved currently. Summary of the Invention

[0012] In view of this, it is necessary to provide a method for extracting and recognizing maritime ship targets from optical remote sensing images.

[0013] The present invention provides a method for extracting and recognizing marine ship targets from optical remote sensing images. The method comprises the following steps: a. Input a visible light remote sensing image; b. Introduce covariance statistics and homologous similarity metrics, and use a visual saliency method based on covariance joint features to globally detect the sea surface in the remote sensing image, generating a single-scale saliency map; c. Downsample the generated single-scale saliency map to establish a multi-scale saliency map, and obtain the final saliency map through a multiplicative fusion mechanism and normalization fusion; d. According to the gray-scale statistical features of the obtained final saliency map, calculate the detection threshold of the marine ship targets on the sea surface, binarize the saliency map, achieve rough segmentation of the saliency map, and mark it back to the original remote sensing image to find the regions of each target and separate suspected targets from the sea surface background; e. Establish a training set and a test set. The training set includes positive samples and negative samples, and the test set includes positive samples and negative samples; f. Design CF-Fourier features, embed aggregated channel features (ACF) and pyramid features (FGPM), and establish a framework to obtain feature vectors that can be used for training and testing; g. According to the established framework and the positive and negative samples in the training set, train a model, and use a Boosting decision tree to classify the feature vectors in step f; h. According to the established framework and the positive and negative samples in the test set, test the candidate regions to complete the extraction and recognition of marine ship targets from optical remote sensing images.

[0014] Preferably, step a includes:

[0015] Input an optical remote sensing image f(x, y) with a spatial resolution of H×W. There are ships, sea fog, thick clouds, islands, etc. in the remote sensing image. Among them, the sizes and color polarities of the ships are different, and their positions on the sea surface are randomly distributed.

[0016] Preferably, step b includes:

[0017] Step S21: According to the input remote sensing image, calculate the brightness of pixel m, extract gradient features in the horizontal and vertical directions, second-order derivative features of brightness, and features of brightness L, opponent color dimensions a, and b in the Lab color space close to human vision, and form a nine-dimensional feature vector f with the position coordinates (x, y). m :

[0018]

[0019] Step S22: Divide the remote sensing image into square regions R of the same size, calculate the feature mean, and symmetrically construct a 9×9 covariance feature matrix as the region descriptor S with f. m :

[0020]

[0021] Step S23: Perform Cholesky decomposition on the region descriptor to obtain each row vector L in the upper triangular matrix i , then the region descriptor is equivalent to a set of points S in Euclidean space:

[0022]

[0023] Combine the feature mean μ and the point set S to obtain C R Encode the feature vector ψ with the computing ability of Euclidean space μ (C R ):

[0024] ψ μ (C R ) = (μ, s 1 , s 2 ,..., s k ,, s k+1 ,.s k+2 .., s 2k );

[0025] Step S24: Through context similarity measurement, find the significance of the T most similar measurement representation regions. The formula is as follows:

[0026]

[0027]

[0028] Step S25: Design a homologous similarity weight function w j on the basis of the above significance to enhance the contrast and obtain a sparse map of the significant region:

[0029]

[0030] Among them, the weight function uses the inverse function of feature distance to measure and is defined as a Gaussian function:

[0031]

[0032]

[0033] Preferably, the step d includes:

[0034] Adopt the OTSU method to obtain an adaptive segmentation threshold T to establish a connected region for extracting the target:

[0035]

[0036] Preferably, the step f specifically includes:

[0037] Step S61: The gradient of the planar image I(x, y) at the pixel (x, y) is represented as (D(x, y), θ(D(x, y))), and the continuous gradient direction pulse curve is calculated as:

[0038] h(ζ) = ||D(x, y)||δ(ζ - θ(D(x, y)));

[0039] Step S62: Use Fourier analysis for the gradient direction pulse curve:

[0040]

[0041] Coefficient

[0042] Step S63: Rotate the image within the vector field and find the conditions for rotational invariance and design the self - guiding kernel function P j (r):

[0043]

[0044]

[0045]

[0046]

[0047] Step S64: Adopt the above kernel function for convolution modeling. According to the above conditions for rotational invariance, the Fourier HOG rotational invariance descriptor is expressed as follows:

[0048]

[0049] Step S65: Introduce the circular frequency filter CF. Utilize the brightness difference between the ship and the surrounding background to design the gray - value change pattern of the ship target, and calculate the discrete Fourier transform DFT of the gray value at the pixel (i, j):

[0050]

[0051] Step S66: The extracted rotational invariance gradient features and circular frequency features are fed into the classifier to distinguish whether it is a real ship or a false alarm. Aiming at the slow efficiency of directly collecting image pyramid features, on the basis of the basic scale d 0 , estimate the scale factor λ to achieve fast pyramid feature estimation for different scales d 1 :

[0052] F d1 = F d0 ·(d 0 / d 1 )-λ 。

[0053] Preferably, step g specifically includes the following steps:

[0054] By taking positive and negative samples in a ratio of 1:3 as the input end of the model, training the model, and classifying and generating the confidence scores of candidate regions, and using the intersection over union as the criterion for determining whether it is a true target.

[0055] Preferably, step h specifically includes:

[0056] Testing the suspected ship target area fragments extracted in step d to determine whether it is a true target or a false alarm. If it is a true target, it is retained; if it is a false alarm, it is eliminated, and finally marked in the input image.

[0057] This application does not have many complex parameter settings and does not rely on prior knowledge of the sea surface background and target distribution characteristics. Aiming at the characteristics of ship targets under the sea surface background, a method combining visual saliency detection based on covariance joint statistical features is proposed. The covariance estimation of the region is combined with homologous similarity weighted fusion to correct the deficiencies, enhancing the overall advantage, thereby suppressing the interference of the sea surface background. The multi-scale multiplication fusion enhances the overall continuity of the detected targets and the distinguishability between targets, and efficiently searches for the sea surface target area. For false alarms such as thick clouds and islands that may appear in the image, the aggregated channel features - feature pyramid acceleration framework (ACF-FPGM) embedded with CF-Fourier spatial frequency domain joint features is used to further identify the detected targets to determine whether the detected targets are ships, greatly reducing the false alarm rate and improving the detection accuracy.

[0058] In addition, the detection and identification time of this application is in seconds, with good real-time performance. There is an obvious improvement in the degree of automation, which can realize the rapid discovery, positioning, and quantity determination of ship targets under a large range of sea areas and multiple background interferences, and has good detection robustness. It lays a foundation for further calculating intelligence information such as the position and heading of each ship by combining with the attitude data of the unmanned aerial vehicle platform or satellite, as well as the classification and recognition of ship targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flowchart of the method for extracting and identifying ship targets in optical remote sensing images of the present invention;

[0060] Figure 2 It is a flowchart of the method for extracting and identifying ship targets in optical remote sensing images provided by an embodiment of the present invention;

[0061] Figure 3 It is a schematic diagram of the visual saliency target extraction process provided by an embodiment of the present invention;

[0062] Figure 4 Schematic diagram of multi-scale fusion effect diagram provided by the embodiment of the present invention: Among them, Figure 4 (a) Original image; Figure 4 (b) σ = 2 -4 ; Figure 4 (c) σ = 2 -5 ; Figure 4 (d) σ = 2 -6 ; Figure 4 (e) Fusion saliency map;

[0063] Figure 5(a) is the original image I provided by the embodiment of the present invention;

[0064] Figure 5(b) is the directional gradient map Dx / Dy provided by the embodiment of the present invention;

[0065] Figure 5(c) is a schematic diagram of the Fourier analysis coefficients of the gradient provided by the embodiment of the present invention:

[0066] Figure 6 It is the CF feature gray value change pattern and feature schematic diagram provided by the embodiment of the present invention: Among them, Figure 6 (a) Ships; Figure 6 (b) Gray scale statistical chart; Figure 6 (c) Pseudo-color map;

[0067] Figure 7 It is a schematic diagram of the fine discrimination process provided by the embodiment of the present invention. Detailed implementation manners

[0068] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0069] Refer to Figure 1 、 Figure 2 As shown, it is the operation flow chart of the preferred embodiment of the method for extracting and recognizing marine ship targets from optical remote sensing images of the present invention.

[0070] Step S1, input the visible light remote sensing image f(x, y) with a spatial resolution of H*W. Specifically:

[0071] Input the optical remote sensing image f(x, y) with a spatial resolution of H×W. There are ships, sea fog, heavy clouds, islands, etc. in the remote sensing image. Among them, the sizes and color polarities of the ships are different, and their position distributions on the sea surface are also very random.

[0072] Step S2, introduce the covariance statistic and the homology similarity measure, and use the visual saliency method based on the covariance joint feature to globally detect the sea surface in the remote sensing image to generate a single-scale saliency map. Specifically:

[0073] Step S21: According to the input remote sensing image, calculate the brightness of pixel m, extract the gradient features in the horizontal and vertical directions, the second-order derivative feature of brightness, and the brightness L, opponent color dimensions a, and b features in the Lab color space close to human vision. The above three types of features respectively correspond to Figure 3 the three-row image sequence in the second column of the figure, and form a nine-dimensional feature vector f with the position coordinates (x, y) m :

[0074]

[0075] Step S22: Divide the remote sensing image into square regions R of the same size, calculate the feature mean, and symmetrically construct a 9×9 covariance feature matrix as the region descriptor S with f m :

[0076]

[0077] Step S23: Perform Cholesky decomposition on the region descriptor to obtain each row vector L in the upper triangular matrix i , then the region descriptor is equivalent to a set of point sets S in the Euclidean space:

[0078]

[0079] Merge the feature mean μ and the point set S to obtain C R Encode the feature vector ψ with the computing ability in the Euclidean space μ (C R ):

[0080] ψ μ (C R ) = (μ, s 1 , s 2 ,..., s k ,, s k+1 ,. s k+2 .., s 2k )

[0081] Step S24: Through context similarity measurement, use the Euclidean distance to represent this similarity, and find the T most similar measurements to represent the significance of the region. Among them, the context includes: the radius is three times the length of the division unit, that is, R i The number range is 1-9 and does not include the number 5 of R. In this embodiment, T = 5, and the formula is as follows:

[0082]

[0083]

[0084] Step S25: Based on the above saliency, a homologous similarity weight function w is designed j , enhance the contrast, and obtain a sparse map of the salient region:

[0085]

[0086] Among them, the weight function uses the inverse function of feature distance to measure and is defined as a Gaussian function:

[0087]

[0088]

[0089] Step S3, downsample the generated single-scale salient map to establish a multi-scale salient map, and obtain the final salient map through a multiplicative fusion mechanism and normalization fusion. Specifically:

[0090] According to the above single-scale salient map generation process, extend it to multiple scales, balance the adversarial relationship between regional representation ability and the spatial resolution of the salient map, adopt a multi-scale product fusion strategy and normalize it. Please refer to Figure 4 , in this embodiment, 3 scales are used, Γ = {σ|2 k}(k = -4, -5, -6), as shown in Figure 4 (b)- Figure 4 (d), the final salient map ( Figure 4 (e)) has a better effect of removing cloud and fog at fine scales and is beneficial to highlighting ship targets at coarse scales:

[0091]

[0092] Step S4, according to the gray-scale statistical characteristics of the obtained final salient map, calculate the detection threshold of the sea surface ship target, binarize the salient map, realize the rough segmentation of the salient map, and mark it back to the original remote sensing image to find the area of each target and separate suspected targets and the sea surface background. Specifically:

[0093] In this embodiment, the OTSU method is adopted to obtain an adaptive segmentation threshold T to establish a connected region for extracting targets:

[0094]

[0095] Step S5, please refer to Figure 7 together to establish a training set and a test set. The training set includes positive samples and negative samples; the test set includes positive samples and negative samples.

[0096] In this embodiment, a training set and a test set are established. The dataset contains a total of 630 images, and each image has a size of 56 pixels * 56 pixels. The positive samples include various ships under different backgrounds, and the size of the ships ranges from 6 to 20 pixels; the negative samples come from background interferences that may exist at sea, such as sea waves, wake waves, thin clouds, thick cloud masses, islands, etc.

[0097] Step S6: Design CF-Fourier features, embed aggregated channel features (ACF) and pyramid features (FGPM), and establish a framework to obtain feature vectors that can be used for training and testing. Specifically:

[0098] Step S61: As shown in Figures 5(a) and 5(b), the gradient at pixel (x, y) of the planar image I(x, y) is represented as (D(x, y), θ(D(x, y))), then the continuous gradient direction impulse curve is calculated as:

[0099] h(ζ) = ||D(x, y)||δ(ζ - θ(D(x, y)))

[0100] Step S62: Use Fourier analysis on the gradient direction impulse curve:

[0101]

[0102] Coefficient The corresponding Fourier domain coefficient image is shown in Figure 5(c);

[0103] Step S63: Rotate the image within the vector field and find the conditions for rotational invariance and design the self-guided kernel function P j (r):

[0104]

[0105]

[0106]

[0107]

[0108] Step S64: Use the above kernel function for convolution modeling. According to the above conditions for rotational invariance, the Fourier HOG rotational invariance descriptor is expressed as follows:

[0109]

[0110] Step S65: Introduce a circular frequency filter (CF), utilize the brightness difference between the ship and the surrounding background, and design the gray value change pattern of the ship target. Please refer to Figure 6, it is necessary to calculate the discrete Fourier transform (DFT) of the gray value at pixel (i, j):

[0111]

[0112] Step S66: The extracted rotation-invariant gradient features and circular frequency features will be fed into a classifier to distinguish whether it is a real ship or a false alarm. Aggregate channel features (ACF) are used to refine the features structurally. The obtained ACF is input into a boosting decision tree, where fast detection rate and low computational requirements are crucial. Aiming at the slow efficiency of directly collecting image pyramid features, on the basis of the basic scale d 0 , the fast pyramid feature estimation of different scales d 1 is achieved by estimating the scale factor λ:

[0113] F d1 = F d0 ·(d 0 / d 1 ) -λ

[0114] Step S7, according to the established framework and the positive and negative samples in the training set, train the model, and use the Boosting decision tree to classify the feature vectors in Step S6.

[0115] By taking the positive and negative samples in a ratio of 1:3 as the input end of the model, train the model, and classify to generate the confidence score of the candidate region. Use the intersection over union (iou) as the criterion for determining whether it is a real target.

[0116] Step S8, according to the established framework and the positive and negative samples in the test set, test the candidate regions to complete the extraction and recognition of marine ship targets in optical remote sensing images.

[0117] In this embodiment, the suspected ship target area fragments extracted in Step S4 are set to 56*56 and fed into the model for testing to determine whether it is a real target or a false alarm. If it is a real target, it is retained; if it is a false alarm, it is eliminated and finally marked in the input image.

[0118] This application includes visual saliency segmentation extraction and fine discrimination of the supervision system. In the visual saliency segmentation extraction stage, by constructing the covariance features belonging to the sea ship target, using the homologous similarity to design the weights of the second-order statistics, the single saliency map obtained after the optimal similarity measurement can well weaken the background and highlight the target; finally, a multi-scale fusion strategy and an adaptive threshold segmentation module are designed to achieve efficient unsupervised extraction of the sea ship area under multiple environments such as different target scales, chaotic cloud and fog backgrounds, sea clutter and wake wave interference. In the fine discrimination stage of the supervision system, the ship spatial frequency domain CF-Fourier HOG feature is designed, which has rotational invariance. Under the framework of aggregated channel features and fast feature pyramid acceleration, the identification of the ship target and the elimination of false alarms are completed.

[0119] Although the present invention has been described with reference to the current preferred embodiments, those skilled in the art should understand that the above preferred embodiments are only used to illustrate the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle scope of the present invention shall be included within the scope of the present invention's rights protection.

Claims

1. A method for extracting and recognizing maritime ship targets from optical remote sensing images, characterized in that, the method comprises the following steps: a. Input visible light remote sensing images; b. Introduce covariance statistics and homologous similarity metrics, and use the visual saliency method based on covariance joint features to globally detect the sea surface in the remote sensing images to generate a single-scale saliency map; c. Downsample the generated single-scale saliency map to establish a multi-scale saliency map, and obtain the final saliency map through a multiplicative fusion mechanism and normalization fusion; d. According to the gray-scale statistical features of the obtained final saliency map, calculate the detection threshold of the maritime ship targets, binarize the saliency map, achieve rough segmentation of the saliency map, and mark it back to the original remote sensing image, find the regions of each target, and separate suspected targets and the sea surface background; e. Establish a training set and a test set, where the training set includes positive samples and negative samples, and the test set includes positive samples and negative samples; f. Design CF-Fourier features, embed the aggregated channel features ACF and the pyramid features FGPM, and establish a framework to obtain feature vectors that can be used for training and testing; g. According to the established framework and the positive and negative samples in the training set, train the model, and use the Boosting decision tree to classify the feature vectors in step f; h. According to the established framework and the positive and negative samples in the test set, test the candidate regions to complete the extraction and recognition of maritime ship targets from optical remote sensing images; The step b includes: Step S21: Calculate the brightness of pixel m based on the input remote sensing image, extract the gradient features in the horizontal and vertical directions, the second-order derivative feature of brightness, and the brightness L in the Lab color space close to human vision m , the opponent color dimension a m and b m features, and form a nine-dimensional feature vector f with the position coordinates (x, y) m : Step S22: Divide the remote sensing image into square regions R of the same size, calculate the feature mean value, and use f m Symmetrically construct a 9×9 covariance feature matrix as the region descriptor S: Step S23: Perform Cholesky decomposition on the region descriptor to obtain each row vector L in the upper triangular matrix i , then the region descriptor is equivalent to a set of points S in Euclidean space: Combine the feature mean μ and the point set S to obtain C R Encode the feature vector ψ with the computing ability in Euclidean space μ (C R ): ψ μ (C R ) = (μ, s 1 , s 2 ,..., s k , s k+1 , s k+2 ..., s 2k ); Step S24: Through context similarity metrics, find the most similar T metric representation regions for saliency, and the formula is as follows: Step S25: Based on the above saliency, a homologous similarity weight function w is designed j , enhance the contrast to obtain a sparse map of the salient region: where the weight function uses the feature distance inverse function to measure and is defined as a Gaussian function: The step c includes: According to the generation process of the single-scale saliency map, extend it to multiple scales, balance the adversarial relationship between the regional representation ability and the spatial resolution of the saliency map, and adopt a multi-scale product fusion strategy and normalization to obtain the final saliency map; The step f specifically includes: Step S61: The gradient at pixel (x, y) of the planar image I(x, y) is expressed as (D(x, y), θ(D(x, y))), and the continuous gradient direction impulse curve is calculated as: h(ζ) = D(x, y)δ(ζ - θ(D(x, y))); Step S62: Use Fourier analysis on the gradient direction impulse curve: Coefficient Step S63: Rotate the image within the vector field and find the conditions for rotational invariance and the designed self-guiding kernel function P j (r): Step S64: Adopt the above kernel function convolution modeling. According to the above conditions of rotational invariance, the Fourier HOG rotational invariance descriptor is expressed as follows: Step S65: Introduce a circular frequency filter CF, utilize the brightness difference between the ship and the surrounding background, design the gray-scale value change pattern of the ship target, and calculate the discrete Fourier transform DFT of the gray-scale value at pixel (i, j): Step S66: The extracted rotation-invariant gradient features and circumferential frequency features are fed into a classifier to distinguish whether it is a real ship or a false alarm. Regarding the slow efficiency of directly collecting image pyramid features, on the basis of the basic scale d 0 , the fast pyramid feature estimation of different scales d 1 is achieved by estimating the scale factor λ: F d1 = F d0 ·(d 0 d 1 ) -λ 。 2. The method according to claim 1, characterized in that, the step a includes: Input an optical remote sensing image f(x, y) with a spatial resolution of H×W. There are ships, sea fog, thick clouds, and islands in the remote sensing image. Among them, the sizes and color polarities of the ships are different, and their positions on the sea surface are randomly distributed.

3. The method according to claim 2, characterized in that, Step d includes the following: Using the OTSU method to obtain an adaptive segmentation threshold T to establish connected regions for target extraction:

4. The method according to claim 3, characterized in that, Step g specifically includes the following steps: By taking the positive and negative samples in a ratio of 1:3 as the input end of the model, training the model, and classifying to generate the confidence score of the candidate region, using the intersection over union as the determination criterion for whether it is a true target.

5. The method according to claim 4, characterized in that, Step h specifically includes: Testing the suspected ship target region fragments extracted in step d to determine whether they are real targets or false alarms. Retain those that are real targets and eliminate those that are false alarms, and finally mark them in the input image.

Images