Structural information guided SAR (Synthetic Aperture Radar) target fine granularity identification method and system
Through the SAR target fine-grained identification method guided by structural information, using technologies such as YOLOv5 and DCN modules, the problem of insufficient robustness and generalization in SAR image detection is solved, and more accurate fine-grained classification and model identification is achieved, reducing the problem of low robustness and generalization in SAR image detection, thereby achieving more accurate fine-grained classification and model identification. The false alarm rate.
Patent Information
- Application Number
- CN202410861508.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-28
- Publication Date
- 2024-11-12
AI Technical Summary
In the prior art, the SAR image detection method has weak robustness and generalization, high false alarm rate, serious missed detection and error detection, and it is difficult to accurately classify fine-grained size, especially for specific models of targets such as ships. superior.
Using the SAR target fine-grained recognition method guided by structure information, through the combination of YOLOv5 data preprocessing, DCN and attention module, the deformable convolution and improved IoU calculation method are used to add structural information detection branches to optimize the identification network to improve the matching of the The recognition ability of small and medium-sized targets.
It significantly improves the accuracy and robustness of fine-grained recognition of SAR targets, reduces false alarm rates, enhances the ability to distinguish different types of targets, and improves the overall performance of the identification system.
Smart Images

Figure CN118941914A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of computer vision and image processing, and in particular to a method and system for fine-grained recognition of SAR targets guided by structural information. Background Art
[0002] Image detection is the first step of the SAR automatic target recognition (ATR) system, and its execution effect has a profound impact on subsequent recognition tasks. The traditional SAR image detection method is to model the geometric features, grayscale statistical features and texture features of the target, and mainly rely on manually designed templates to distinguish the target from the background, and the recognition rate for specific scenes is relatively high. The Constant False Alarm Rate (CFAR) algorithm based on clutter statistics and threshold extraction is a classic traditional algorithm. This algorithm separates the background and the target by modeling the background noise. Although CFAR reduces the detection loss rate, it is dependent on the assumed distribution of background clutter, has weak resistance to strong clutter interference, and requires artificial setting of the appropriate detection window size. Therefore, the robustness and generalization are weak, and a reliable prediction model cannot be established. In addition, limited by the accuracy of the sea and land segmentation processing results, the traditional detection algorithm has a high false alarm rate and serious missed detection and false detection in the near-coast and port area scenes, which brings great difficulties to the deployment and application of the recognition system.
[0003] With the rapid development of deep learning technology, its application in the field of SAR target detection has achieved remarkable results. Deep learning models can self-learn complex features in data, eliminating the traditional image preprocessing steps, and can output prediction results with probabilistic information. In 2017, Google and CMU's research emphasized the importance of large-scale datasets to the success of deep learning, pointing out that with the increase of training data, the performance of computer vision tasks will also increase linearly. Despite this, the breadth and diversity of SAR image datasets are still insufficient, which limits the further development of deep learning in this field. To solve this problem, the first SAR dataset SSDD specifically for ship detection was released in 2017, followed by SSDD+, SAR-Ship, HRSID, AIR-SARShip1.0 / 2.0, SAR-AIRcraft-1.0 and FUSAR-Ship, etc., which provide a solid data foundation for the research and application of SAR image detection. The introduction of deep learning has ushered in a new era for SAR image detection, and its performance has surpassed traditional SAR target detection methods. The existing deep network model is trained and inferred in an end-to-end manner, which can automatically extract multi-scale and multi-resolution features, realize real-time and accurate target detection in complex environments, and bring a revolutionary new approach to the target detection task of SAR images.
[0004] Information about key component areas is crucial for the application of SAR images. For example, in the analysis of ship and aircraft targets, information about the head and tail can assist in determining the heading; the aspect ratio of the target helps to identify different types of targets within the same category. In search and rescue missions, information about the status of key components (such as radar and power units) is crucial because they can help analyze the current status of the rescue target. In the field of military reconnaissance, by analyzing the structural distribution of the target's key components, the specific model and behavioral intention of the target can be identified, the overall battlefield situation can be controlled, and correct decision-making and deployment can be made.
[0005] However, research on fine-grained recognition, that is, distinguishing and identifying targets at a more microscopic level, is still insufficient, and this limitation affects the performance of real-time monitoring applications based on SAR systems. The scarcity of fine-grained recognition research is partly due to the fact that publicly available fine-grained datasets of SAR image targets usually only mark targets such as aircraft and ships as foregrounds, without detailed classification by specific model. In addition, the intra-class differences of targets in SAR images are small, which increases the difficulty of fine-grained recognition. For example, in aircraft target recognition, aircraft of different models may show highly similar scattering characteristics, making them difficult to distinguish. To improve recognition accuracy, research needs to consider introducing additional information to assist the recognition process, such as the size, shape, structural characteristics and other related attributes of the aircraft, to achieve more accurate fine-grained classification. With the improvement of the resolution of SAR imaging systems, the detailed information in the target structures such as aircraft and ships can be clearly recorded, which improves the distinction, recognition and detection of targets of different categories and scales. In the field of SAR image analysis, researchers are gradually turning to explore the scattering characteristics of targets. This shift marks the deepening of SAR image interpretation. Sun Xian et al. proposed a scattering association classifier based on the significant differences in the discrete distribution of spatial structures of different aircraft models in SAR images. The classifier quantifies the discrete degree of modeled aircraft targets and dynamically adjusts the weights of sample pairs using the differences in the discrete distribution of targets, thereby guiding the deep learning network to learn more discriminative inter-class feature representations. This method not only enhances the network's attention to the key component areas of the aircraft, but also effectively reduces the interference of background noise.
[0006] In summary, the existing technologies have technical problems such as weak robustness and generalization, high false alarm rate, missed detection and false detection, and difficulty in accurate fine-grained classification of specific target models. Summary of the invention
[0007] The technical problem to be solved by the present invention is: how to solve the technical problems in the prior art of weak robustness and generalization, high false alarm rate, missed detection and wrong detection, and difficulty in accurate fine-grained classification of specific target models.
[0008] The present invention adopts the following technical solutions to solve the above technical problems: The SAR target fine-grained recognition method guided by structural information includes:
[0009] S1. Collect SAR target data and use the YOLOv5 strategy to preprocess the SAR target image to obtain preprocessed SAR target data;
[0010] S2, sending the pre-processed SAR target data to the preset DCN module and the preset attention module to perform structural information guidance operations on key features of the pre-processed SAR target data to obtain guided enhanced data;
[0011] S3, sending the guided enhancement data to the detection head of the structure information guided recognition network to generate a suitable recognition result; in the network detect stage, using no less than 2 parallel branches to receive the input feature map, respectively performing target category regression operation, key point detection operation and target structure information prediction to obtain a suitable recognition result;
[0012] S4. Use Focaler-SloU for network training, take Focaler-SloU as the IoU improvement strategy, add structural information to guide the recognition network, calculate the Focaler-loU loss for applicable recognition results, and compare it with the category regression constraints. According to the key point detection constraints and the target structure information prediction constraints, perform parameter optimization guidance operations in the back propagation process.
[0013] The present invention conducts research on fine-grained recognition and pose estimation of SAR targets under supervised training, constructs the structural information of the target through scattering key points, and further guides the network to optimize the recognition results. On the basis of adding structural information, the present invention optimizes the recognition network in a targeted manner to adapt to the image characteristics of SAR targets, specifically adding a deformable convolution module DCN, an attention module, and an improved IoU calculation method.
[0014] In a more specific technical solution, in S1, the enhancement stage of the YOLOv5 strategy includes: Mosaic enhancement, Mixup, random perspective, and HSV augmentation.
[0015] In a more specific technical solution, in S2, in the structure information guided operation, the head part of the YOLOv8m-pose network is adjusted, and a structure information detection branch is added based on the target detection box regression component and the key point detection component, and the number of color child channels is the structure information number Nstruc.
[0016] The present invention selects YOLOv8m-pose as the baseline network, and identifies a new method YOLOv8-improved for fine-grained target recognition in SAR images based on structural information guidance. The method can effectively use key points to form structural information to guide the network for fine-grained recognition.
[0017] In a more specific technical solution, in S2, in the preset DCN module and the preset attention module, the offset result obtained by convolution of the input feature map using deformable convolution is used to perform convolution learning on each pixel of the input feature map to obtain a set of offsets (x, y), and the same prediction offset is used on different channels of the same input feature map.
[0018] In a more specific technical solution, S2 includes:
[0019] S21. Using ordinary convolution, sample a set of pixels from the input feature map, and calculate the pixels through the convolution operation to obtain the fractional offset:
[0020]
[0021] S22, using deformable convolution to modify the pixel points to perform convolution kernel deformation operation, wherein Δp is used n Represents p n To expand:
[0022] {Δp n |n=1,2,……,N},
[0023] The deformable convolution is calculated using the following logic:
[0024]
[0025] S23, in the preset DCN module, performing bilinear interpolation sampling on the decimal offset, obtaining and using the applicable offset, performing an offset operation, and extracting SAR image features;
[0026] S24, sending the SAR image features to a preset attention module for fusion operation, wherein the preset attention module includes: a channel attention module and a spatial attention module.
[0027] In a more specific technical solution, S24 also includes:
[0028] S241. In the channel attention module, for the input feature map, the maximum pooling operation and the average pooling operation are used to aggregate the spatial information, and the output is a C-dimensional pooled feature map: F avg 、F max ;
[0029] S242, pool the C-dimensional feature map F avg 、F max Input the multilayer perceptron MLP with hidden layer to obtain the channel attention map. In the multilayer perceptron MLP with hidden layer, the number of input information neurons is compressed on the hidden layer neurons, and the channel attention map is added to obtain the channel attention map Mc;
[0030] S243, in the spatial attention module, the input feature map is subjected to average pooling and maximum pooling operations according to the channel dimension to obtain a concatenated feature map, and a spatial attention map M is generated by using a preset size convolutional layer. s .
[0031] The present invention selects structural information with clear physical meanings such as line segment ratio, structural angle, direction vector and aspect ratio and adds them to the fine-grained recognition network. The feature information contained in SAR images is not rich, and it is difficult for the network to extract high-dimensional features with strong characterization ability, so certain auxiliary information is needed to guide the network to perform fine-grained recognition. The scattering key points contain the positional spatial information of the key structures and components of the target. By using the positional relationship between these key points, the target structure information features with identifiable characteristics can be formed.
[0032] In a more specific technical solution, in S242, the channel attention map Mc is determined using the following logic:
[0033]
[0034] Among them, F avg and F max is the C-dimensional pooling feature map.
[0035] The present invention uses a deformable convolutional module DCN (Deformable Convolutional Network) to replace the detection layers P3 and P4 in the YOLOv8m-pose backbone network to enhance the feature representation of small objects. This modification can adaptively adjust the receptive field of the network during model sampling to improve the model's attention to small and medium-sized targets. After adding the deformable convolution module, because the receptive field of the location of the small target has adaptively changed, when the prediction box is regressed, the model can better adjust the regression parameters of the prediction box, so that the model increases its attention to small targets, thereby improving the overall performance of the model.
[0036] In a more specific technical solution, in S243, the following logic is used to obtain the spatial attention map M s :
[0037]
[0038] The present invention adds an attention module to the neck end of YOLOv8m-pose to adaptively select and adjust the channel weight and spatial weight of the feature map, so as to better capture and represent the important features of the SAR image target. The attention module combines the spatial module and the channel module, and uses the global average pooling and global maximum pooling pooling strategies, so as to effectively prevent information loss and thus achieve better results.
[0039] In a more specific technical solution, S4 includes:
[0040] S41. Define Focaler-SloU using the following logic:
[0041]
[0042] Among them, IoU focaler is the reconstructed Focaler-loU, loU is the original loU value, and IoU is adjusted by adjusting the values of d and u. focaler Focus on different regression samples;
[0043] S42. Determine the Focaler-loU loss using the following logic:
[0044] L Focaler-IoU =1-IoU focaler
[0045] S43, using the following logic, the Focaler-loU loss is applied to the following loU-based bounding box regression loss function L Focaler-DIoU , L Focaler-CIoU , L Focaler-EIoU and L Focaler-SIoU :
[0046] L Focaler-DIoU =L DIoU +IoU-IoU Focaler
[0047] L Focaler-CIoU =L CIoU +IoU-IoU Focaler
[0048] L Focaler-EIoU =L EIoU +IoU-IoU Focaler
[0049] L Focaler-SIoU =L SIoU +IoU-IoU Focaler .
[0050] In a more specific technical solution, the structure information guided SAR target fine-grained recognition system includes:
[0051] SAR data preprocessing module, used to collect SAR target data, and use YOLOv5 strategy to preprocess the SAR target image to obtain preprocessed SAR target data;
[0052] A structure information guidance module is used to send the pre-processed SAR target data to the preset DCN module and the preset attention module to perform a structure information guidance operation on the key features of the pre-processed SAR target data to obtain guidance enhancement data. The structure information guidance module is connected to the SAR data pre-processing module;
[0053] The network recognition module is used to send the guided enhancement data to the detection head of the structure information guided recognition network to generate a suitable recognition result; in the network detect stage, at least two parallel branches are used to receive the input feature map, and target category regression operation, key point detection operation and target structure information prediction are performed respectively to obtain a suitable recognition result. The network recognition module is connected to the structure information guided module;
[0054] The parameter optimization guidance module is used to use Focaler-SloU for network training, take Focaler-SloU as the IoU improvement strategy, add structural information to guide the recognition network, calculate the Focaler-loU loss for applicable recognition results, and compare it with the category regression constraint. According to the key point detection constraint and the target structure information prediction constraint, the parameter optimization guidance operation is performed in the back propagation process. The parameter optimization guidance module is connected to the network recognition module.
[0055] Compared with the prior art, the present invention has the following advantages:
[0056] The present invention conducts research on fine-grained recognition and pose estimation of SAR targets under supervised training, constructs the structural information of the target through scattering key points, and further guides the network to optimize the recognition results. On the basis of adding structural information, the present invention optimizes the recognition network in a targeted manner to adapt to the image characteristics of SAR targets, specifically adding a deformable convolution module DCN, an attention module, and an improved IoU calculation method.
[0057] The present invention selects YOLOv8m-pose as the baseline network, and identifies a new method YOLOv8-improved for fine-grained target recognition in SAR images based on structural information guidance. The method can effectively use key points to form structural information to guide the network for fine-grained recognition.
[0058] The present invention selects structural information with clear physical meanings such as line segment ratio, structural angle, direction vector and aspect ratio and adds them to the fine-grained recognition network. The feature information contained in SAR images is not rich, and it is difficult for the network to extract high-dimensional features with strong characterization ability, so certain auxiliary information is needed to guide the network to perform fine-grained recognition. The scattering key points contain the positional spatial information of the key structures and components of the target. By using the positional relationship between these key points, the target structure information features with identifiable characteristics can be formed.
[0059] The present invention uses a deformable convolutional module DCN (Deformable Convolutional Network) to replace the detection layers P3 and P4 in the YOLOv8m-pose backbone network to enhance the feature representation of small objects. This modification can adaptively adjust the receptive field of the network during model sampling to improve the model's attention to small and medium-sized targets. After adding the deformable convolution module, because the receptive field of the location of the small target has adaptively changed, when the prediction box is regressed, the model can better adjust the regression parameters of the prediction box, so that the model increases its attention to small targets, thereby improving the overall performance of the model.
[0060] The present invention adds an attention module to the neck end of YOLOv8m-pose to adaptively select and adjust the channel weight and spatial weight of the feature map, so as to better capture and represent the important features of the SAR image target. The attention module combines the spatial module and the channel module, and uses the global average pooling and global maximum pooling pooling strategies, so as to effectively prevent information loss and thus achieve better results.
[0061] The present invention solves the technical problems existing in the prior art of weak robustness and generalization, high false alarm rate, missed detection and wrong detection, and difficulty in accurate fine-grained classification of specific target models. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 A schematic diagram of the basic steps of the structural information-guided SAR target fine-grained recognition method according to Embodiment 1 of the present invention;
[0063] Figure 2 This is a schematic diagram of the YOLOv8m-pose+DCN+attention network structure of Example 1 of the present invention;
[0064] Figure 3 Schematic diagram of the attention structure of Embodiment 1 of the present invention;
[0065] Figure 4 This is a schematic diagram of the Yolov8-improved SAR image fine-grained recognition network guided by structural information according to Example 2 of the present invention. DETAILED DESCRIPTION
[0066] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0067] Example 1
[0068] like Figure 1 As shown, the structural information-guided SAR target fine-grained recognition method provided by the present invention includes the following specific steps:
[0069] Step S1, data preprocessing;
[0070] In the data preprocessing process of this embodiment, the YOLOv5 strategy is adopted. During training, the enhancement stage of the above strategy includes but is not limited to: Mosaic enhancement, Mixup enhancement, random perspective and HSV augmentation.
[0071] Step S2: Send the preprocessed data to the DCN and attention module to guide and enhance the key features in the image;
[0072] like Figure 2 As shown, in this embodiment, the backbone network structure of YOLOv8 uses a layer of 3x3 convolution with a step size of 2 to downsample the feature map, followed by a C2f module to further enhance the features therein, and the basic depth parameters of C2f can be set to, for example, 3 / 6 / 9 / 3 respectively.
[0073] In this embodiment, the structural information guidance is modified in the head part of the YOLOv8m-pose network. A branch of structural information detection is added on the basis of the original target detection box regression and key point detection. The number of channels is the number of structural information Nstruc, and the value of Nstruc depends on the number of channels required for the structural information.
[0074] In the process of the DCN and attention mechanism of this embodiment, the deformable convolution is the result of the convolution on the input feature map to obtain the offset. A set of offsets (x, y) is learned through convolution at each pixel point of the input feature map, and the same predicted offset is used on different channels of the same feature map.
[0075] In this embodiment, the calculation process of ordinary convolution is to sample a group of pixels from the input feature map and then use the convolution operation to calculate the sampled results to obtain the result:
[0076]
[0077] In this embodiment, for deformable convolution, the sampling result is modified to achieve the effect of deforming the convolution kernel. n Represents p n Expand, where:
[0078] {Δpn |n=1,2,……,N}, the following formula shows the calculation process of deformable convolution:
[0079]
[0080] In this embodiment, the offset obtained by the above convolution operation is a decimal and cannot be directly sampled and used. Therefore, DCN adopts a bilinear interpolation sampling method to achieve the effect of using the offset. The above formula describes the process of bilinear interpolation.
[0081] like Figure 3 As shown, in this embodiment, after extracting features, the features are sent to the attention module for fusion. In this embodiment, the attention mechanism includes but is not limited to: a channel module and a spatial module; in this embodiment, a channel attention module can be constructed by using a channel module and a spatial module in a parallel manner, for example.
[0082] In the operation process of the channel attention module of this embodiment, the maximum pooling operation and the average pooling operation are used to aggregate the spatial information of the input feature map, and the output is two C-dimensional pooled feature maps: F avg and F max ; F avg and F max The input is put into the multilayer perceptron MLP containing the hidden layer, and two channel attention maps with parameters of 1x1xC are obtained. In this embodiment, in order to reduce the number of parameters, the number of neurons in the hidden layer that need to compress the input information is: C / r, where r is called the compression ratio; the channel attention map Mc is obtained by adding the output of the previous step.
[0083] In this embodiment, the aforementioned channel attention map Mc can be expressed by the following formula:
[0084]
[0085] In this embodiment, during the spatial attention operation, F' is firstly subjected to average pooling and maximum pooling operations according to the channel dimension, and two feature maps Fag and Fm are output. The output parameters are 1×H×W, and the two feature maps output are spliced in the channel dimension to obtain the spliced spatial feature map. For the spliced feature map, a convolutional layer with a size of 7x7 is used to further generate the spatial attention map M. s .
[0086] In this embodiment, the spatial attention map M s , which can be expressed by the following formula:
[0087]
[0088] Step S3: The features are sent to the detection head to generate the final recognition result;
[0089] In this embodiment, in the detect stage of the network, three parallel branches receive the same feature map, and perform target category regression, key point detection, and target structure information prediction respectively, and finally use the result of target category regression as the final recognition result.
[0090] Step S4: Use Focaler-SloU to perform network training.
[0091] In this embodiment, in the object detection task, Intersection over Union (IoU) is widely used to weigh the similarity between the prediction result of the model and the actual annotation of the object.
[0092] In this embodiment, the above-mentioned intersection over union ratio is calculated by calculating the intersection and union of the model prediction result and the real annotation area; the intersection is divided by the union to obtain a ratio value between 0 and 1. The larger the IoU value, the higher the similarity between the model prediction result and the real annotation. If the IoU value is greater than the preset threshold, it is considered that the predicted box matches the real box successfully.
[0093] In this embodiment, the present invention selects Focaler-SloU as an IoU improvement strategy, adds it to the recognition network guided by structural information, and calculates the Focaler-loU loss for the output results of the three branches in the above step S3. Compared with the simple category regression constraint, the key point detection and target structure information prediction are also used as constraints, which guides the optimization of parameters during back propagation.
[0094] In this embodiment, Focaler-SloU is defined as follows:
[0095]
[0096] Among them, IoU focaler is the reconstructed Focaler-loU, loU is the original loU value, where [d,u]∈[0,1]. By adjusting the values of d and u, IoU focaler Focus on different regression samples.
[0097] In this embodiment, the Focaler-loU loss is defined as follows:
[0098] L Focaler-IoU =1-IoU focaler
[0099] In this embodiment, the Focaler-loU loss is applied to the following loU-based bounding box regression loss function LFocaler-DIoU , L Focaler-CIoU , L Focaler-EIoU and L Focaler-SIoU :
[0100] L Focaler-DIoU =L DIoU +IoU-IoU Focaler
[0101] L Focaler-CIoU =L CIoU +IoU-IoU Focaler
[0102] L Focaler-EIoU =L EIoU +IoU-IoU Focaler
[0103] L Focaler-SIoU =L SIoU +IoU-IoU Focaler .
[0104] Example 2
[0105] like Figure 4 As shown, in this embodiment, the input of the present invention is a SAR image. The structural information guidance method proposed in the present invention can use the key points and structural information in the image as optimization constraints to guide the fine-grained recognition of SAR targets.
[0106] In the first stage of this embodiment, the method proposed in the present invention adds a deformable convolutional network DCN to the visual layer to enhance feature representation, aiming to improve the adaptability of the convolutional neural network to irregularly shaped target objects and expand its receptive field. Traditional convolution operations can only extract features from regular rectangular receptive fields, but in real scenes, the shapes of many target objects are irregular, such as faces, cars, animals, etc. At this time, deformable convolution is needed to identify these irregularly shaped target objects. Deformable convolution can adaptively adjust the shape and size of the receptive field according to the irregular shape of the target object, thereby improving the robustness of CNN.
[0107] In the second stage of this embodiment, due to the special bottom-up and top-down structure of the neck end, features of different scales are fully integrated here, laying the foundation for subsequent predictions, so the network structure of the neck end greatly affects the algorithm performance. To this end, the present invention uses an attention module to adaptively select and adjust the channel weights and spatial weights of the feature map, so as to better capture and represent the important features of the SAR image target.
[0108] In the third stage of this embodiment, three parallel branches are used to perform target category regression, key point detection and target structure information prediction respectively, and the Focaler-loU loss is calculated for the output results of the three branches to constrain the model training direction.
[0109] In this embodiment, different structural information is added to verify its effectiveness, and the experimental results of each structural information are statistically analyzed. The structural information is divided into angle Angle, vector Vector, segment length ratio SR (Segmentlengthratio) and integrated structure information IS (Integrated structure): Angle represents the angle information composed of any three points among the five key points, for example, the angle value formed by the second key point as the vertex, the first key point and the third key point; Vector represents the vector information composed of any two points among the five key points, such as the unit vector from the first key point to the second key point; SR represents the length ratio information of the line segment composed of several points composed of the five key points, for example, taking points 1, 2, and 3 to obtain the length of three line segments, and the ratio of each line segment to the total length; Integrated structure information IS represents the combination of the above three structural information. After experimental verification, the proposed structural information has obvious effectiveness in guiding the network.
[0110] In summary, this paper proposes a new SAR image fine-grained recognition network Yolov8-improved based on structural information guidance to effectively integrate structural information into the backbone network YOLOv8m-pose to guide the network to perform fine-grained recognition. The selection of structural information is crucial to network performance. Appropriate structural information as a guide can significantly improve recognition accuracy. In addition, the DCN module, attention module, and Focaler-loU have played a significant role in improving recall and precision. The optimized recognition network has shown effective performance improvement in fine-grained recognition tasks.
[0111] The present invention conducts experiments on the SAR5points dataset, adds DCN, attention module, IoU improvement method and structural information, and the optimized recognition network improves AP50 by 2.2% and AP50:95 by 3.3% compared with the original YOLOv8m-pose model. At the same time, the precision and recall rates are also improved to a certain extent, which shows that not only the missed detection of the model is improved, but also the recognition effect of difficult-to-identify targets is better.
[0112] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A SAR target fine-grained recognition method guided by structural information, characterized in that: The method comprises: S1. Collect SAR target data, and use the YOLOv5 strategy to perform data preprocessing on the SAR target image to obtain preprocessed SAR target data; S2, sending the pre-processed SAR target data to a preset DCN module and a preset attention module to perform a structural information guidance operation on key features of the pre-processed SAR target data to obtain guided enhanced data; S3, sending the guided enhancement data to the detection head of the structure information guided recognition network to generate an applicable recognition result; in the network detect stage, using no less than 2 parallel branches to receive the input feature map, respectively performing target category regression operation, key point detection operation and target structure information prediction to obtain the applicable recognition result; S4. Use Focaler-SloU to train the network, use the Focaler-SloU as the IoU improvement strategy, add the structural information to guide the recognition network, calculate the Focaler-loU loss for the applicable recognition result, and compare it with the category regression constraint. According to the key point detection constraint and the target structure information prediction constraint, perform parameter optimization guidance operation in the back propagation process.
2. The structure information guided SAR target fine-grained recognition method according to claim 1, characterized in that: In S1, the enhancement stage of the YOLOv5 strategy includes: Mosaic enhancement, Mixup enhancement, random perspective and HSV augmentation.
3. The structure information guided SAR target fine-grained recognition method according to claim 1, characterized in that: In S2, in the structure information guided operation, the head part of the YOLOv8m-pose network is adjusted, and a structure information detection branch is added based on the target detection frame regression component and the key point detection component, and the number of color child channels is the structure information number Nstruc.
4. The structure information guided SAR target fine-grained recognition method according to claim 1, characterized in that: In S2, in the preset DCN module and the preset attention module, the offset result obtained by convolution of the input feature map using deformable convolution is used to perform convolution learning on each pixel of the input feature map to obtain a set of offsets (x, y), and the same predicted offset is used on different channels of the same input feature map.
5. The structure information guided SAR target fine-grained recognition method according to claim 1, characterized in that: The S2 includes: S21. Using ordinary convolution, sample a group of pixels from the input feature map, and calculate the pixels through convolution operation to obtain a fractional offset: S22, using deformable convolution to modify the pixel point to perform a convolution kernel deformation operation, wherein Δp n Represents p n To expand: {Δp n |n=1,2,……,N}, The deformable convolution is calculated using the following logic: S23, in the preset DCN module, performing bilinear interpolation sampling on the decimal offset, obtaining and using the applicable offset, performing an offset operation, and extracting SAR image features; S24, sending the SAR image features to the preset attention module for fusion operation, wherein the preset attention module includes: a channel attention module and a spatial attention module.
6. The structure information guided SAR target fine-grained recognition method according to claim 5, characterized in that: The S24 further includes: S241. In the channel attention module, for the input feature map, the maximum pooling operation and the average pooling operation are used to aggregate the spatial information, and the output is a C-dimensional pooled feature map: F avg 、F max ; S242, the C-dimensional pooling feature map F avg 、F max Inputting into a multilayer perceptron MLP with a hidden layer to obtain a channel attention map, in the multilayer perceptron MLP with a hidden layer, performing a compression operation on the number of input information neurons on the hidden layer neurons, performing an addition operation on the channel attention map to obtain the channel attention map Mc; S243: In the spatial attention module, average pooling and maximum pooling operations are performed on the input feature map according to the channel dimension to obtain a concatenated feature map, and a spatial attention map M is generated by using a preset size convolutional layer. s .
7. The method for fine-grained SAR target recognition guided by structural information according to claim 6, characterized in that: In S242, the channel attention map Mc is determined using the following logic:
8. The method for fine-grained SAR target recognition guided by structural information according to claim 6, characterized in that: In S243, the following logic is used to obtain the spatial attention map M s : 。 9. The structure information guided SAR target fine-grained recognition method according to claim 1, characterized in that: The S4 includes: S41. Define Focaler-SloU using the following logic: Where, IoU focaler is the reconstructed Focaler-loU, loU is the original loU value, and IoU is adjusted by adjusting the values of d and u. focaler Focus on different regression samples; S42, using the following logic to determine the Focaler-loU loss: L Focaler-IoU =1-IoU focaler S43, using the following logic, the Focaler-loU loss is applied to the following loU-based bounding box regression loss function L Focaler-DIoU , L Focaler-CIoU , L Focaler-EIoU and L Focaler-SIoU : L Focaler-DIoU =L DIoU +IoU-IoU Focaler L Focaler-CIoU =L CIoU +IoU-IoU Focaler L Focaler-EIoU =L EIoU +IoU-IoU Focaler L Focaler-SIoU =L SIoU +IoU-IoU Focaler .
10. A SAR target fine-grained recognition system guided by structural information, characterized in that: The system comprises: A SAR data preprocessing module is used to collect SAR target data and perform data preprocessing on the SAR target image using the YOLOv5 strategy to obtain preprocessed SAR target data; A structure information guiding module, used for sending the pre-processed SAR target data to a preset DCN module and a preset attention module, so as to perform a structure information guiding operation on key features of the pre-processed SAR target data to obtain guided enhanced data, wherein the structure information guiding module is connected to the SAR data pre-processing module; A network recognition module is used to send the guided enhancement data to the detection head of the structure information guided recognition network to generate an applicable recognition result; in the network detect stage, at least two parallel branches are used to receive the input feature map, and target category regression operation, key point detection operation and target structure information prediction are performed respectively to obtain the applicable recognition result, and the network recognition module is connected to the structure information guided module; A parameter optimization guidance module is used to adopt Focaler-SloU for network training, take the Focaler-SloU as the IoU improvement strategy, add the structural information to guide the recognition network, calculate the Focaler-loU loss for the applicable recognition result, compare it with the category regression constraint, perform parameter optimization guidance operation in the back propagation process according to the key point detection constraint and the target structure information prediction constraint, and the parameter optimization guidance module is connected to the network recognition module.
Citation Information
Patent Citations
Small sample SAR target detection and identification method based on meta learning and metric learning
CN116664823A
Traffic target identification method based on improved YOLOv8 model
CN117237915A
Improved YOLOv8-based SAR image ship target detection method, medium and device
CN117689870A
Method and apparatus for end-to-end SAR image recognition, and storage medium
US20210003697A1
Contextual visual-based SAR target detection method and apparatus, and storage medium
US20230184927A1
Cited By
Solenopsis invicta detection method based on improved YOLOv5s
CN119741733A
