Cascade type foresight sonar small target detection method and system and medium

Abstract

The invention provides a cascaded foresight sonar small target detection method and system and a medium. The method comprises the following steps: firstly, downsampling a sonar image; a cross-shaped constant false alarm detector is adopted to preliminarily screen target candidate areas; the same target is merged through clustering of adjacent detection points; extracting the target scale, the signal-to-noise ratio and the area feature for preliminary filtering; a positive and negative sample data set is constructed by using a constant false alarm detection result, and training and classification are performed based on a residual network (ResNet34). According to the method, through multistage progressive processing, false signals are effectively filtered, the accuracy and robustness of small target detection are remarkably improved, meanwhile, the dependence of deep learning on a large amount of manual annotation data is reduced, and the method is suitable for target intelligent detection of forward-looking sonar in a complex marine environment.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic signal processing technology, specifically to a cascaded forward-looking sonar small target detection method, system, and medium, and in particular to a cascaded forward-looking sonar small target detection method that integrates constant false alarm rate (CFAR) detection and residual network. Background Technology

[0002] Forward-looking sonar, as a core sensor for underwater observation, acquires acoustic images that form the basis for underwater target detection, identification, and situational awareness. However, due to the complex and ever-changing underwater propagation environment, various types of ocean clutter, and the inherent characteristics of the sensor itself, sonar images typically exhibit low signal-to-noise ratios, low contrast, and strong background noise, making the reliable detection of small targets a long-standing technical challenge.

[0003] Target detection primarily addresses the question of whether a target exists. Common underwater target detection algorithms include traditional methods such as frame difference and constant false alarm rate (CFAR) detection, as well as deep learning methods such as the YOLO series. Due to the complexity of the underwater environment, small targets have few effective pixels and weak echoes in sonar images. Traditional detection methods are prone to high false alarm rates and high missed detection rates in complex environments, affecting subsequent target decisions. Deep learning detection methods, on the other hand, require a huge amount of labeled data, are prone to overlooking weak real targets at low signal-to-noise ratios, and have high computational complexity.

[0004] Therefore, there is an urgent need to develop a new detection framework that can organically integrate the robustness and interpretability of traditional signal processing methods with the powerful feature learning capabilities of deep learning, so as to achieve high-precision, high-reliability, and low-false-alarm detection of weak targets in forward-looking sonar images under real-world conditions of limited data and complex environments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a cascaded forward-looking sonar small target detection method, system, and medium.

[0006] A cascaded forward-looking sonar small target detection method according to the present invention includes the following steps: Step S1: Downsample the forward-looking sonar image; Step S2: Use a cross-shaped constant false alarm detector to perform preliminary target detection on the downsampled image to obtain a binarized detection image containing candidate targets; Step S3: Cluster the detection points in the binarized detection image to merge adjacent detection points belonging to the same target and generate candidate regions; Step S4: For each candidate region, extract the target scale, signal-to-noise ratio, and region area features, and perform preliminary screening of the candidate regions based on the preset feature thresholds to obtain an optimized candidate region set; Step S5: Based on the selected candidate region set and its experimental annotation information, construct a training dataset containing positive and negative samples, and preprocess the sample images. Step S6: Train a residual network classification model using the training dataset; Step S7: For the forward-looking sonar image to be detected, the optimized candidate region set is obtained by sequentially going through steps S1 to S4. Each candidate region is processed into data in the same format as the training samples and input into the residual network classification model trained in step S6 to obtain the final target classification result.

[0007] Preferably, in step S2, the cross-shaped constant false alarm detector takes the unit under test as the center, selects training units in its row direction and column direction respectively to estimate the background noise, and adaptively calculates the detection threshold based on the estimation result to determine whether the unit under test is a candidate target.

[0008] Preferably, the clustering process in step S3 uses a sliding window to statistically analyze the density of detection points and merges adjacent regions with a density exceeding a set threshold into a single candidate region.

[0009] Preferably, in step S4, the target scale is determined by searching the number of pixels in the original sonar image corresponding to the candidate region along the longitudinal direction of the energy center whose intensity values ​​fall within a specified range.

[0010] Preferably, in step S5, the negative sample is a candidate region generated through steps S1 to S3 but which, after experimental verification, does not contain the real target.

[0011] Preferably, the preprocessing in step S5 includes image normalization, size uniform transformation to preset specifications, and data augmentation operations.

[0012] Preferably, in step S6, the residual network is a deep residual neural network.

[0013] According to the present invention, a cascaded forward-looking sonar small target detection system integrating constant false alarm rate (CFAR) detection and residual network is provided, comprising the following sub-modules: Module M1: Downsamples the forward-looking sonar image; Module M2: Uses a cross-shaped constant false alarm detector to perform preliminary target detection on the downsampled image, and obtains a binarized detection image containing candidate targets; Module M3: Performs clustering processing on the detection points in the binarized detection image to merge adjacent detection points belonging to the same target and generate candidate regions; Module M4: For each candidate region, extract the target scale, signal-to-noise ratio, and region area features, and perform preliminary screening of the candidate regions based on preset feature thresholds to obtain an optimized candidate region set; Module M5: Based on the selected candidate region set and its experimental annotation information, a training dataset containing positive and negative samples is constructed, and the sample images are preprocessed. Module M6: Train a residual network classification model using the training dataset; Module M7: For the forward-looking sonar image to be detected, modules M1 to M4 are executed sequentially to obtain its optimized candidate region set. Each candidate region is processed into data in the same format as the training samples and input into the residual network classification model in module M6 to obtain the final target classification result.

[0014] According to the present invention, a computer-readable storage medium is provided thereon storing a computer program that, when executed by a processor, implements the steps of the cascaded forward-looking sonar small target detection method that integrates constant false alarm rate detection and residual network.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention discloses a cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual networks, which addresses the problems of high false alarm rate and low recognition accuracy in target detection under complex environments in existing technologies. Compared with conventional CFAR algorithms, this algorithm can learn complex, hierarchical feature representations from target data, resulting in better detection capabilities. Compared with deep learning methods, this method can generate negative sample bounding boxes through CFAR detection during residual network training, reducing the cost of manual data annotation.

[0016] 2. This invention achieves accurate detection of small targets by forward-looking sonar through multi-level progressive processing. This method combines the advantages of traditional signal processing and deep learning, achieving complementary strengths. The residual network can learn complex and nonlinear features that constant false alarm rate (CFAR) detection cannot handle, while the CFAR algorithm provides physically meaningful preprocessing, directly filtering out a large amount of obvious non-target background, providing a "cleaner" input to the neural network, allowing the residual network to focus on learning more subtle and discriminative features. Through a multi-layered filtering mechanism, this method filters out the vast majority of false signals, improving the model's generalization ability in unknown environments and effectively reducing the false alarm rate. Attached Figure Description

[0017] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 These are images acquired by a forward-looking sonar test lake in an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow in an embodiment of the present invention; Figure 3 This is a schematic diagram of the cross-shaped constant false alarm detection window in an embodiment of the present invention; Figure 4 This is the result of the cross-shaped constant false alarm rate detection in the embodiments of the present invention; Figure 5 This is the result of clustering the detection points in the embodiments of the present invention; Figure 6 This refers to the target annotation results of the dataset in this embodiment of the invention; Figure 7 This represents the target classification accuracy result in this embodiment of the invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0019] This invention proposes a cascaded forward-looking sonar small target identification method that integrates constant false alarm rate (CFAR) detection and residual networks. This method combines the robustness of traditional algorithms with the strong representational capabilities of deep learning models through a cascaded, coarse-grained processing flow. Through a multi-layered filtering mechanism, it filters out the vast majority of false signals, reducing the false alarm rate and improving identification efficiency and accuracy. This method is applicable to complex marine environments and has practical significance for forward-looking sonar target detection.

[0020] This embodiment uses the measured data from a forward-looking sonar lake test as an example. Typical images of the test data are shown below. Figure 1 As shown, the target used in the experiment was a 0.7 × 0.7 m target sphere. A schematic diagram of the method flow in this embodiment is shown below. Figure 2 As shown, the steps include: (1) Downsampling Sonar images typically have a small number of beams, while the data at each sampling point is tens of times larger than the number of beams. To improve target detection efficiency, sonar images are downsampled in the range direction, as shown in the following formula:

[0021] in, This represents the downsampled sonar image. , These represent the horizontal and vertical coordinates of the downsampled sonar image, respectively. , This represents the number of sampling points after downsampling. , These represent the number of sampling points and the number of beams before downsampling, respectively. This represents the downsampling factor.

[0022] In a preferred embodiment, the sonar image has a beam number of 256 and a distance sampling point number of 8108, therefore the downsampling factor n is 16 times.

[0023] (2) Cross-shaped constant false alarm target detection The reference cells of the cross-shaped constant false alarm rate (CFAR) detector are distributed in the rows and columns of the cells under test, forming a cross. By moving a sliding window in the sonar image, it is possible to analyze whether each data point (cell under test) contains a target. Figure 3 As shown. To prevent strong targets near the unit under test from affecting noise estimation, protective units are set around the unit under test. These protective units do not participate in noise estimation. The size of the crosshair window is set according to the target characteristics of the forward-looking sonar. The local estimation of the row and column is taken as the unit mean to obtain the background estimation result. In a preferred embodiment, considering the existence of side lobes in the sonar image, 10 protective units and 12 training units are used in the row direction, and 4 protective units and 12 training units are used in the column direction. The calculation formula for background estimation near the columns and rows (training units) of the unit under test is as follows:

[0024]

[0025] in, , These represent the background estimation in the row and column directions of the unit under test, respectively. , These represent the intensity values ​​of the training units in the row and column directions of the unit under test, respectively. , These represent the number of training units in the row and column directions, respectively.

[0026] Based on the calculated background estimation, the detection threshold formula is as follows:

[0027] in, To detect threshold, This is the threshold coefficient. The detection threshold adapts to the data and can adaptively adjust the detection threshold as the detection window slides. Choosing an appropriate threshold coefficient can keep the false alarm probability at a constant level. To prevent missed detections, the detection threshold can be set to be more lenient.

[0028] When the intensity value of the unit to be tested is greater than the detection threshold, it indicates the presence of the target. The unit value is marked as 1, and the rest of the area is marked as 0, resulting in the image. , refer to Figure 4As shown, this process generates multiple candidate regions.

[0029] (3) Clustering of neighboring unit detection points Constant false alarm rate (CFAR) detection can remove most of the background noise and clutter in sonar images. Due to the complexity of the underwater environment, the brightness of the same target may be uneven in the sonar image, leading to fragmented detection points. Dividing the detection area into grid cells and counting the number of detection points within a sliding window (window size can be set to 5*7) allows clustering of detection points within adjacent cells, merging similar objects spatially. The clustering result is shown below. Figure 5 As shown. The calculation for each pixel is performed as follows:

[0030] in, , These represent the row and column lengths of the downsampled image, respectively. The image obtained after clustering. , The row and column values ​​of the sliding coordinates. and The sliding window is half the width of the row and column, and the window size is [value missing]. , This is the clustering threshold, and its value is determined based on the target characteristics.

[0031] (4) Preliminary feature extraction and filtering In the clustered image, the pixel with the highest intensity in the candidate region is taken as the target energy center, and the target scale, signal-to-noise ratio, and area features of each candidate region are calculated.

[0032] The target scale feature calculation formula is as follows: Search both sides of the target energy center in the longitudinal direction of each candidate region, considering the difference between the intensity value and the energy center to be within a threshold. (For example The number of units within the range is the target's vertical scale. Based on the calculated target scale, regions that do not meet the target characteristics are filtered out, such as regions with a target scale of less than 3 pixels.

[0033]

[0034] in, Represents the vertical search range. These represent the x and y coordinates of the target energy center in each candidate region, respectively. The x-coordinate represents the search range at the target scale.

[0035] The signal-to-noise ratio of the rectangular panes surrounding each candidate region is calculated sequentially, and the difference between the intensity value within the pane and the energy center is within a threshold value. (For example: The number of units within the specified range represents the number of signal units, and the remainder represents the number of noise units. Based on the target characteristics, regions with excessively low signal-to-noise ratios are filtered out. The target signal-to-noise ratio calculation formula is as follows:

[0036]

[0037] in, and Set the half-width of both rows and columns of the matrix window to 20 pixels. and These represent the number of signal units and noise units, respectively.

[0038] After clustering, the pixel count of candidate regions in the image is used as the region area. A threshold is set to retain regions whose area matches the target characteristics, as shown in the following formula:

[0039] in, This represents the lower limit of the region area, for example, set to 10. By extracting features to filter candidate regions, unsuitable candidate regions can be eliminated, resulting in an optimized set of candidate regions and improving the efficiency of subsequent processing.

[0040] (5) Creation of deep learning datasets Of the 97 underwater target images acquired by forward-looking sonar, the actual target locations in the experiment were marked using an annotation tool, referring to... Figure 6 As shown, the dataset is divided into a training set and a test set in an 8:2 ratio. The labeled target information in the training set is the positive sample. The target candidate region is generated in the training dataset through steps (1)-(3). The candidate region that does not contain the real target is labeled as the negative sample. A total of 97 positive samples and 2442 negative samples are obtained. The bounding box of the same size is drawn for each sample with the maximum intensity of the candidate region as the center, and the corresponding category label is labeled.

[0041] (6) Image preprocessing The sample images in the dataset are linearly normalized to map pixel values ​​to the [0,1] interval, as shown in the following formula:

[0042]

[0043] in, The pixel unit representing the sample. Represents the maximum value of the sample pixels. Represents the minimum value of the sample pixels. This represents the normalization result. , Represents the size of the sample rows and columns.

[0044] The target size is uniformly transformed to 224×224 pixels to adapt to ResNet34 input.

[0045] A data augmentation method with horizontal flipping (probability 0.5) is applied to the dataset to increase the model's generalization ability.

[0046] (7) Training the residual network A ResNet34 network was trained using a constructed forward-looking sonar small target dataset. Input samples were converted from grayscale images to RGB images to adapt them to the network input. A cross-entropy loss function and the Adam optimizer were used. The network was trained iteratively with a learning rate of 0.001, a batch size of 16, and a training iteration count of 0.0001 for 20 iterations. The recognition accuracy on the training set was 99.7%.

[0047] (8) Validation of test set results The test dataset is processed through steps (1)-(3) to detect candidate regions. Boundary boxes of the same size are drawn for each sample, centered on the maximum intensity of the candidate region, to serve as test samples. After normalization and scaling of the samples, they are input into the trained residual network to extract deep features and obtain the detection results. The classification accuracy of the test set is 99.8%, as shown below. Figure 7 As shown, the target was detected correctly 100% of the time in the test set, with only one non-target region being incorrectly detected as a target.

[0048] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0049] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual networks, characterized in that, Includes the following steps: Step S1: Downsample the forward-looking sonar image; Step S2: Use a cross-shaped constant false alarm detector to perform preliminary target detection on the downsampled image to obtain a binarized detection image containing candidate targets; Step S3: Cluster the detection points in the binarized detection image to merge adjacent detection points belonging to the same target and generate candidate regions; Step S4: For each candidate region, extract the target scale, signal-to-noise ratio, and region area features, and perform preliminary screening of the candidate regions based on the preset feature thresholds to obtain an optimized candidate region set; Step S5: Based on the selected candidate region set and its experimental annotation information, construct a training dataset containing positive and negative samples, and preprocess the sample images. Step S6: Train a residual network classification model using the training dataset; Step S7: For the forward-looking sonar image to be detected, the optimized candidate region set is obtained by sequentially going through steps S1 to S4. Each candidate region is processed into data in the same format as the training samples and input into the residual network classification model trained in step S6 to obtain the final target classification result.

2. The cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual network as described in claim 1, characterized in that, In step S2, the cross-shaped constant false alarm detector takes the unit under test as the center, selects training units in its row direction and column direction respectively to estimate the background noise, and adaptively calculates the detection threshold based on the estimation result to determine whether the unit under test is a candidate target.

3. The cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual network as described in claim 1, characterized in that, The clustering process described in step S3 uses a sliding window to statistically analyze the density of detection points and merges adjacent regions with a density exceeding a set threshold into a single candidate region.

4. The cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual network as described in claim 1, characterized in that, In step S4, the target scale is determined by searching the number of pixels in the original sonar image corresponding to the candidate region along the longitudinal direction of the energy center whose intensity values ​​fall within a specified range.

5. The cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual network as described in claim 1, characterized in that, In step S5, the negative sample is a candidate region generated through steps S1 to S3 but which, as verified by experiments, does not contain the real target.

6. The cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual network as described in claim 1, characterized in that, The preprocessing described in step S5 includes image normalization, size uniform transformation to preset specifications, and data augmentation operations.

7. The cascaded forward-looking sonar small target detection method integrating constant false alarm rate (CFAR) detection and residual network as described in claim 1, characterized in that, In step S6, the residual network is a deep residual neural network.

8. A cascaded forward-looking sonar small target detection system integrating constant false alarm rate (CFAR) detection and residual network, characterized in that, Includes the following sub-modules: Module M1: Downsamples the forward-looking sonar image; Module M2: Uses a cross-shaped constant false alarm detector to perform preliminary target detection on the downsampled image, and obtains a binarized detection image containing candidate targets; Module M3: Performs clustering processing on the detection points in the binarized detection image to merge adjacent detection points belonging to the same target and generate candidate regions; Module M4: For each candidate region, extract the target scale, signal-to-noise ratio, and region area features, and perform preliminary screening of the candidate regions based on preset feature thresholds to obtain an optimized candidate region set; Module M5: Based on the selected candidate region set and its experimental annotation information, a training dataset containing positive and negative samples is constructed, and the sample images are preprocessed. Module M6: Train a residual network classification model using the training dataset; Module M7: For the forward-looking sonar image to be detected, modules M1 to M4 are executed sequentially to obtain its optimized candidate region set. Each candidate region is processed into data in the same format as the training samples and input into the residual network classification model in module M6 to obtain the final target classification result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the cascaded forward-looking sonar small target detection method that integrates constant false alarm rate detection and residual network as described in any one of claims 1 to 7.

Images