A SAR image ship wake detection method based on multiple image entropy features

Abstract

The present application relates to the technical field of target detection, and particularly relates to a SAR image ship wake detection method based on multiple image entropy features, comprising: acquiring a SAR sea surface image, and obtaining a to-be-detected SAR image block through preprocessing; based on each to-be-detected SAR image block, entropy features are extracted and spliced respectively to obtain corresponding entropy feature vectors; the obtained entropy feature vectors are input into a trained classifier for classification to obtain corresponding classification results as detection results of the corresponding SAR image blocks. The present application can realize high-accuracy detection of SAR image ship wake targets.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method for detecting ship wakes in SAR images based on multiple image entropy features. Background Technology

[0002] SAR (Surveillance-Resonance) is an active microwave sounding imaging system. Due to its ability to observe ground features at high resolution and long distances, both day and night, it has become one of the most important remote sensing tools for ocean observation. High-resolution SAR images are beneficial for acquiring information on maritime targets; therefore, marine target detection using SAR images has become an important research topic. Ship wakes are the tracks left on the rough sea surface by moving ships disturbing the surface. Compared to the ship itself, wakes appear larger in SAR images, are easier to detect, and can be used to deduce ship speed and actual position. Therefore, conducting effective research on SAR image wake detection is of great significance for marine environmental protection and monitoring.

[0003] SAR (Shock and Reflection) generates SAR images by transmitting microwave signals and receiving backscattered electromagnetic echoes from detected objects. Therefore, the representation of any object in a SAR image is related to its background environment and its own electromagnetic scattering characteristics. When radar observes the sea, it inevitably receives scattered echo signals from the ocean surface, i.e., sea clutter. When sea state is high, wakes are obscured by complex sea clutter backgrounds, reducing their observability in SAR images and increasing the difficulty of wake detection.

[0004] The core of the target detection problem is to find the characteristic differences between the target and the background, separate the target from the background, and complete the target detection. Based on the generation mechanism, wakes have a definite linear geometric structure. Therefore, current SAR image ship wake detection methods mainly revolve around single linear shape features, transforming the wake detection problem into a straight line detection problem against a sea clutter background. However, these methods suffer from low accuracy and are prone to missed detections and false detections. Summary of the Invention

[0005] To address the issues of low accuracy and poor reliability in existing SAR image ship wake detection technologies, this invention provides a SAR image ship wake detection method based on multiple image entropy features.

[0006] This invention provides a method for detecting ship wakes in SAR images based on multiple image entropy features, including:

[0007] Acquire SAR sea surface images and, through preprocessing, obtain SAR image patches to be detected;

[0008] Based on each SAR image patch to be detected, entropy features are extracted and concatenated to obtain the corresponding entropy feature vectors. Among them, the entropy features include frequency domain energy spectrum entropy features, power spectrum entropy features, and fractional Fourier domain energy spectrum entropy features.

[0009] The obtained entropy feature vector is input into the trained classifier for classification, and the corresponding classification result is used as the detection result of the corresponding SAR image patch. The classifier is trained with the entropy feature vector corresponding to the sample SAR image patch as input and the corresponding classification result as output. The entropy feature vector corresponding to the sample SAR image patch is obtained by extracting entropy features from the sample SAR image patch and performing feature stitching. The classification result is sea clutter or wake.

[0010] Optionally, the preprocessing includes:

[0011] The acquired SAR sea surface image is divided into non-repeating blocks according to a preset size, and the grayscale intensity of each image block is normalized.

[0012] Optionally, the frequency domain energy spectral entropy features are extracted using the following method:

[0013] Perform a Fourier transform on the SAR image patch to obtain the corresponding spectrum;

[0014] Based on the obtained spectrum, calculate the energy of the SAR image patch in each frequency band in the frequency domain and its proportion relative to the total spectral energy;

[0015] Based on the energy proportion of each frequency band in the frequency domain, the corresponding frequency domain energy spectral entropy characteristics are calculated.

[0016] Optionally, the power spectral entropy features are extracted as follows:

[0017] Based on SAR image patches, determine the corresponding autocorrelation function;

[0018] Based on the determined autocorrelation function, the corresponding power spectrum is obtained through Fourier transform;

[0019] Based on the obtained power spectrum, calculate the energy of the SAR image patch in each frequency band in the power spectrum domain and its proportion relative to the total spectral energy.

[0020] Based on the energy proportion of each frequency band in the power spectral domain, the corresponding power spectral entropy characteristics are calculated.

[0021] Optionally, the fractional Fourier domain energy spectral entropy features are extracted using the following method:

[0022] Determine the orders p1 and p2;

[0023] Based on the determined orders p1 and p2, the corresponding fractional Fourier domain spectrum is obtained through fractional Fourier transform.

[0024] Based on the obtained fractional Fourier domain spectrum, calculate the energy of the SAR image patch in each frequency band in the fractional Fourier domain and its proportion relative to the total spectral energy.

[0025] Based on the energy percentage of each frequency band in the fractional Fourier domain, the corresponding fractional Fourier domain energy spectral entropy characteristics are calculated.

[0026] Optionally, the classifier is trained in the following manner:

[0027] Acquire a set of known sample SAR sea surface images and select the region of interest.

[0028] For the region of interest image, non-overlapping blocks are divided into normalized blocks according to a preset size to obtain sample SAR image blocks with classification results;

[0029] Based on each sample SAR image patch, entropy features are extracted and concatenated to obtain the entropy feature vector corresponding to the sample SAR image patch;

[0030] The entropy feature vectors corresponding to the obtained sample SAR image patches are divided into training set and test set, and the constructed classifier is trained using the training set and test set until the accuracy of the classifier meets the preset requirements.

[0031] Optionally, the classifier is a support vector machine.

[0032] This invention provides a method for detecting ship wakes in SAR images based on multiple image entropy features. Starting from the perspective of image texture differences, this invention describes the characteristics of sea clutter and wake images using multiple image entropy features, which can better characterize the differences between sea clutter and wakes. Combined with a trained classifier, it can effectively improve the accuracy of wake detection in SAR images, thereby solving the problems of high false alarm rates and low accuracy of current detection methods that rely on the linear shape features of wakes. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart of a SAR image ship wake detection method based on multiple image entropy features provided by an embodiment of the present invention;

[0035] Figure 2 It is a SAR image of the sea surface;

[0036] Figure 3 This is a comparison diagram of the energy spectrum entropy of sea clutter and wake;

[0037] Figure 4 This is a comparison of the power spectral entropy of sea clutter and wake;

[0038] Figure 5 This is a comparison of the energy spectral entropy of sea clutter and wake in the fractional Fourier domain;

[0039] Figure 6 It is the precision-recall curve of the trained classifier on the test dataset. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] As mentioned earlier, existing SAR image ship wake detection methods mainly revolve around a single linear shape feature, transforming the wake detection problem into a straight line detection problem against a sea clutter background. However, these methods have two problems: firstly, wakes are weak targets, and due to the influence of sea clutter, their linear features on SAR images are not clear, easily leading to missed detections; secondly, sea clutter may also contain wave waves with linear structures, resulting in excessively high false alarm rates. Therefore, relying solely on a single linear shape feature makes it difficult to guarantee the reliability of the detection results and cannot meet the requirements for high-accuracy wake detection.

[0042] Reliably extracting features that clearly distinguish the target from the sea clutter background is the most crucial step in wake detection. Backscattered signals are related to surface roughness; different intensities of backscattered signals will exhibit different grayscale intensities in the image. Wakes are observable because the disturbance of the ship alters the local roughness, creating ripples and forming a texture different from the sea surface. Sea clutter is a non-stationary process, and its texture is more complex than that of a wake. This invention considers the difference in texture between sea clutter and wakes, i.e., the difference in entropy, and therefore introduces the concept of entropy to distinguish between them. Entropy describes the complexity of the texture; the more complex the texture, the greater the entropy value. In view of this, this invention abandons the existing method that relies on linear wake features and proposes a SAR image ship wake detection method based on multiple image entropy features.

[0043] The following describes the specific implementation of the above concept.

[0044] Please refer to Figure 1 This invention provides a method for detecting ship wakes in SAR images based on multiple image entropy features. The method includes:

[0045] Step 100: Acquire SAR sea surface image and obtain SAR image patch to be detected through preprocessing;

[0046] Step 102: Based on each SAR image patch to be detected, the corresponding entropy feature vector is obtained by extracting entropy features and performing feature concatenation.

[0047] Among them, entropy features include frequency domain energy spectrum entropy features, power spectrum entropy features, and fractional Fourier domain energy spectrum entropy features;

[0048] Step 104: Input the obtained entropy feature vector into the trained classifier for classification, and obtain the corresponding classification result as the detection result of the corresponding SAR image patch;

[0049] The classifier is trained by taking the entropy feature vector corresponding to the sample SAR image patch as input and the corresponding classification result as output.

[0050] The entropy feature vector corresponding to the sample SAR image patch is obtained by extracting entropy features from the sample SAR image patch and then stitching the features together.

[0051] The classification result is marine clutter or wake.

[0052] This invention overcomes the limitations of existing methods that rely solely on a single geometric feature for detection, resulting in a high false alarm probability. It employs a detection method based on multiple image spectral entropy features to fully characterize the differences between wakes and sea clutter in images, thereby reducing the false alarm probability and improving the accuracy of SAR image wake detection.

[0053] The following description Figure 1 The execution method of each step is shown.

[0054] Optionally, for step 100, the preprocessing includes:

[0055] The acquired SAR sea surface image is divided into non-repeating blocks according to a preset size, and the grayscale intensity of each image block is normalized.

[0056] Normalization is performed, which maps the grayscale intensity of the image to the range of 0 to 1. The expression is:

[0057]

[0058] Among them, I i_o This represents the i-th block, where max(·) and min(·) represent finding the maximum and minimum grayscale values, respectively. i This represents the i-th SAR image patch obtained after normalization.

[0059] The above embodiments help eliminate the influence of different image values ​​on texture recognition and facilitate processing. The preset size should be the same as the size of the sample SAR image patch so that it can be input into the machine learning classifier.

[0060] Optionally, for step 102, the frequency domain energy spectral entropy features are extracted in the following manner:

[0061] Perform a Fourier transform on the SAR image patch to obtain the corresponding spectrum, i.e., F(u,v), where F(·) represents the two-dimensional discrete Fourier transform, u and v represent frequency domain parameters, u ranges from 1 to U, and v ranges from 1 to V; U and V represent the maximum values ​​of frequency domain parameters u and v, respectively.

[0062] Based on the obtained spectrum, the energy of the SAR image patch in each frequency band in the frequency domain and its proportion relative to the total spectral energy are calculated; that is, the energy in each frequency band is statistically analyzed as E(u,v)=|F(u,v)|. 2 Calculate the ratio p of energy in each frequency band to the total spectral energy. u,v The expression is:

[0063]

[0064] Based on the energy proportion of each frequency band in the frequency domain, that is, the proportion of energy in each frequency band relative to the total spectral energy, the corresponding frequency domain energy spectral entropy characteristic is calculated, and the expression is:

[0065]

[0066] Sea clutter is a non-stationary process with complex frequency components, while the wake has a more concentrated frequency. By using Fourier transform, the spectrum of the image is obtained, and the energy of each frequency band and the proportion of each frequency band to the total energy are calculated. Based on this, the energy spectral entropy of the image in the frequency domain is extracted to describe the energy difference between sea clutter and wake in the frequency domain. Since the energy of the wake is more concentrated than that of sea clutter, the corresponding entropy value is also smaller than that of sea clutter.

[0067] Furthermore, the power spectral entropy features are extracted as follows:

[0068] Based on SAR image patches, the corresponding autocorrelation function is determined, i.e.

[0069] Based on the determined autocorrelation function, the corresponding power spectrum is obtained through Fourier transform. Where M and N represent the size of the SAR image patch, respectively;

[0070] Based on the obtained power spectrum, the energy of the SAR image patch in each frequency band of the power spectrum domain and its proportion relative to the total spectral energy are calculated; the energy of the SAR image patch in each frequency band of the power spectrum domain can be represented as PSD(u,v); the proportion p of the energy in each frequency band to the total power energy is calculated. e u,v The expression is:

[0071]

[0072] Based on the energy proportion of each frequency band in the power spectral domain, the corresponding power spectral entropy feature is calculated, and the expression is:

[0073]

[0074] Power is also a reflection of energy. The greater the energy, the higher the power. Therefore, the power spectral entropy of sea clutter is greater than that of the wake.

[0075] Furthermore, the fractional Fourier domain energy spectral entropy features are extracted using the following method:

[0076] Determine the orders p1 and p2;

[0077] Based on the determined orders p1 and p2, the corresponding fractional Fourier spectrum is obtained through fractional Fourier transform, expressed as:

[0078]

[0079]

[0080] Among them, I i(m,n) represents the pixel with coordinates (m,n) in the i-th SAR image patch, where m ranges from 1 to M, n ranges from 1 to N, and M and N represent the size of the SAR image patch; K p1,p2 (m,n,u,v) represents the kernel function, α=p1×π / 2 and β=p2×π / 2 refer to the rotation angles of the SAR image patch in two dimensions after passing through the two-dimensional fractional Fourier transform, and j is used to represent the imaginary part;

[0081] Based on the obtained fractional Fourier domain spectrum, the energy E of the SAR image patch in each frequency band in the fractional Fourier domain is calculated. FrFT (u,v) and the percentage p relative to the total spectral energy f u,v The expression is:

[0082]

[0083] Based on the energy proportion of each frequency band in the fractional Fourier domain, the corresponding fractional Fourier domain energy spectral entropy characteristic is calculated, and the expression is:

[0084]

[0085] The orders p1 and p2 can be determined based on the sea state in the SAR sea surface image. p1 and p2 can be the same or different, and their values ​​can range from 0.8 to 1.1. At a suitable fractional order, the energy of the wake is more easily accumulated, resulting in more explicit information and lower entropy. The specific orders p1 and p2 can be determined empirically. The energy spectral entropy of sea clutter in the fractional Fourier domain is greater than that of the wake.

[0086] Using the above method, multiple entropy features related to texture distribution can be quickly identified, and these features can be concatenated to form an entropy feature vector, which can then be used as input to a classifier.

[0087] Alternatively, the classifier can be trained in the following manner:

[0088] Acquire a set of known sample SAR sea surface images and select the region of interest.

[0089] For the region of interest image, non-overlapping blocks are divided into normalized blocks according to a preset size to obtain sample SAR image blocks with classification results;

[0090] Based on each sample SAR image patch, entropy features are extracted and concatenated to obtain the entropy feature vector corresponding to the sample SAR image patch;

[0091] The entropy feature vectors corresponding to the obtained sample SAR image patches are divided into training set and test set, and the constructed classifier is trained using the training set and test set until the accuracy of the classifier meets the preset requirements.

[0092] Known sample SAR sea surface image sets can be obtained by collecting publicly available grayscale SAR ocean image sets, such as the HRSID dataset. The data currently released in this dataset mainly consists of ocean observation data, which involves data from three platforms: RadarSat-2, Sential-1B, and TerraSAR-X, covering four polarization modes and resolutions of 0.5 meters, 1 meter, and 3 meters. Figure 2 A raw SAR sea surface image is shown. Regions of interest (ROIs) are defined as both the pure sea surface and areas with wake targets. ROIs are acquired from the image set and then divided into non-overlapping blocks of size M×N. The classification result for SAR image blocks from the pure sea surface region is sea clutter, and the classification result for SAR image blocks from the wake target region is the wake. Non-overlapping block division helps reduce redundant information during classifier training. The entropy feature vector for each sample SAR image block is determined in the same way as the entropy feature vector for the SAR image block to be detected, ensuring that the trained classifier can input the entropy feature vector and output the corresponding classification result.

[0093] This invention utilizes sample SAR image patches to extract entropy features. One hundred sample SAR image patches of sea clutter and one sample SAR image patch of wake are selected, and appropriate orders p1 = 1.1 and p2 = 1.1 are given. The resulting comparison maps of energy spectrum entropy, power spectrum entropy, and energy spectrum entropy in the fractional Fourier domain for sea clutter and wake are shown below. Figures 3 to 5 As shown, image entropy features can effectively describe the characteristics of sea clutter and wakes.

[0094] Based on the training set, a classifier is trained using machine learning. Machine learning has the advantage of requiring fewer training samples. The classifier transforms the wake detection problem into a binary classification problem in the feature space, where the wake is labeled as the positive class. That is, the trained classifier provides the label corresponding to the input feature vector: +1 indicates a wake, otherwise, a wave surface. In one specific embodiment, a support vector machine (SVM) is chosen as the classifier, offering advantages such as fewer training samples and higher robustness. In other embodiments, other classifiers can be selected as needed.

[0095] Optionally, based on the test set, the performance of the classifier can be tested by plotting a precision-recall curve, where precision and recall are expressed as follows:

[0096]

[0097]

[0098] Here, TP represents the number of positive samples in the test set that were identified as wakes, FP represents the number of negative samples in the test set that were misidentified as wakes, and FN represents the number of positive samples in the test set that were misidentified as sea surface. Whether a sample was misidentified can be determined based on the label corresponding to the SAR image patch. By extracting different entropy features and training a classifier based on these features, wake detection in SAR images can be achieved. Figure 6 The ROC curve of the trained classifier on the test set is shown.

[0099] Optionally, the method further includes:

[0100] Step 106: stitch each SAR image block and its corresponding detection results back into the original SAR sea surface image to display the detection results for each region of the SAR sea surface image.

[0101] Using the above method, the detection results determined by classification can be displayed intuitively on the original SAR sea surface image.

[0102] In summary, this invention provides a SAR image ship wake detection method based on multiple image entropy features. This invention proposes three image entropy features for ship wake detection: image energy spectral entropy in the frequency domain, image power spectral entropy, and image energy spectral entropy in the fractional Fourier domain. Under SAR image conditions of different sea states, polarizations, and resolutions, the proposed three features exhibit strong identifiability. Compared to existing detection methods based on single geometric features, such as Radon transform, the proposed method has superior detection capabilities and can effectively improve the accuracy of SAR image wake detection. This invention overcomes the limitation of existing methods that rely solely on single geometric features, leading to a high false alarm probability. By employing a detection method that combines multiple image spectral entropy features, it fully characterizes the differences between wakes and sea clutter in the image, reducing the false alarm probability and improving the accuracy of SAR image wake detection.

[0103] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting ship wakes in SAR images based on multiple image entropy features, characterized in that, include: Acquire SAR sea surface images and, through preprocessing, obtain SAR image patches to be detected; Based on each SAR image patch to be detected, entropy features are extracted and concatenated to obtain the corresponding entropy feature vectors. Among them, the entropy features include frequency domain energy spectrum entropy features, power spectrum entropy features, and fractional Fourier domain energy spectrum entropy features. The frequency domain energy spectral entropy feature is extracted using the following method: Perform a Fourier transform on the SAR image patch to obtain the corresponding spectrum; Based on the obtained spectrum, calculate the energy of the SAR image patch in each frequency band in the frequency domain and its proportion relative to the total spectral energy; Based on the energy proportion of each frequency band in the frequency domain, calculate the corresponding frequency domain energy spectrum entropy characteristics; The fractional Fourier domain energy spectral entropy features are extracted using the following method: Determine the order and ; Based on a given order and The corresponding fractional Fourier domain spectrum is obtained by performing a fractional Fourier transform. Based on the obtained fractional Fourier domain spectrum, calculate the energy of the SAR image patch in each frequency band in the fractional Fourier domain and its proportion relative to the total spectral energy. Based on the energy proportion of each frequency band in the fractional Fourier domain, the corresponding fractional Fourier domain energy spectrum entropy characteristics are calculated. The obtained entropy feature vector is input into the trained classifier for classification, and the corresponding classification result is used as the detection result of the corresponding SAR image patch. The classifier is trained with the entropy feature vector corresponding to the sample SAR image patch as input and the corresponding classification result as output. The entropy feature vector corresponding to the sample SAR image patch is obtained by extracting entropy features from the sample SAR image patch and performing feature stitching. The classification result is sea clutter or wake.

2. The method according to claim 1, characterized in that, The preprocessing includes: The acquired SAR sea surface image is divided into non-repeating blocks according to a preset size, and the grayscale intensity of each image block is normalized.

3. The method according to claim 1, characterized in that, The power spectral entropy features are extracted using the following method: Based on SAR image patches, determine the corresponding autocorrelation function; Based on the determined autocorrelation function, the corresponding power spectrum is obtained through Fourier transform; Based on the obtained power spectrum, calculate the energy of the SAR image patch in each frequency band in the power spectrum domain and its proportion relative to the total spectral energy. Based on the energy proportion of each frequency band in the power spectral domain, the corresponding power spectral entropy characteristics are calculated.

4. The method according to claim 2, characterized in that, The classifier is trained in the following manner: Acquire a set of known sample SAR sea surface images and select the region of interest. For the region of interest image, non-overlapping blocks are divided into normalized blocks according to a preset size to obtain sample SAR image blocks with classification results; Based on each sample SAR image patch, entropy features are extracted and concatenated to obtain the entropy feature vector corresponding to the sample SAR image patch; The entropy feature vectors corresponding to the obtained sample SAR image patches are divided into training set and test set, and the constructed classifier is trained using the training set and test set until the accuracy of the classifier meets the preset requirements.

5. The method according to claim 1, characterized in that, The classifier is a support vector machine.

Images