SAR (Synthetic Aperture Radar) image target identification method based on interval enhanced center loss function
By introducing an interval-enhanced center loss function, the class discrimination capability of the SAR image target recognition model is improved, solving the problems of large intra-class differences and high inter-class similarity, and achieving higher recognition accuracy.
Patent Information
- Application Number
- CN202511673080.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-24
AI Technical Summary
In SAR image target recognition, there are problems of large intra-class differences and high inter-class similarity. Existing center loss functions are not very adaptable when dealing with scenarios with high inter-class similarity, resulting in poor recognition performance.
An interval-enhanced center loss function is adopted. By calculating the Euclidean distance between the target feature and the class center, an inter-class loss is introduced, which forces a sufficiently large distance between different classes. This loss function is then jointly trained with the cross-entropy loss function to form an interval-enhanced center loss function for SAR image target recognition.
It improved the accuracy of SAR image target recognition to 97.59%, which is 1.03% higher than using the cross-entropy loss function alone, and 0.57% higher than the training results of using the center loss function in conjunction with the cross-entropy loss function.
Smart Images
Figure CN121564583A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of image target recognition and deep learning, and particularly to a SAR image target recognition method based on the interval-enhanced center loss function. Background Technology
[0002] Synthetic Aperture Radar (SAR) is a high-resolution imaging radar that can obtain high-resolution radar images similar to optical photography under weather conditions with extremely low visibility.
[0003] A radar that uses the relative motion between the radar and the target to synthesize a larger equivalent antenna aperture from a smaller actual antenna aperture through data processing is also called a composite aperture radar.
[0004] Synthetic aperture radar (SAR) is characterized by high resolution, all-weather operation, and the ability to effectively identify camouflage and penetrate concealment. The resulting high azimuth resolution is equivalent to that provided by a large-aperture antenna.
[0005] Like most other radars, synthetic aperture radar determines distance by the time difference between transmitting an electromagnetic pulse and receiving the target echo. Its resolution is related to the pulse width or pulse duration; the narrower the pulse width, the higher the resolution.
[0006] Synthetic aperture radar is typically mounted on aircraft or satellites, and is divided into airborne and spaceborne types.
[0007] Synthetic Aperture Radar (SAR) measures range and performs two-dimensional imaging based on the platform's motion trajectory. Its two-dimensional coordinate information consists of range information and azimuth information perpendicular to the range. Azimuth resolution is directly proportional to beamwidth and inversely proportional to antenna size. Just as optical systems require large lenses or mirrors to achieve high precision, radar also requires large antennas or apertures to obtain clear images when operating at low frequencies.
[0008] Because aircraft flight paths are irregular and highly variable, they can cause image defocusing. Inertial and navigation sensors must be used to compensate for antenna motion, and the imaging data must be repeatedly processed to autofocus and generate an image with maximum contrast. Therefore, synthetic aperture radar imaging must operate in a side-looking mode, transmitting coherent signals within a synthetic aperture length, and then coherently processing the received signals to obtain an electronic mosaic.
[0009] The brightness of a radar image pixel is proportional to the energy reflected from the corresponding area on the target. The total is the radar cross-section (RCS), expressed in area. The degree of backscattering is expressed as the normalized radar cross-section (NDR), in decibels (dB). Typical NDRs for the Earth's surface are: brightest +5 dB, darkest -40 dB. Synthetic aperture radar cannot resolve details that the human eye and cameras can distinguish, but its operating wavelength allows it to penetrate clouds and dust.
[0010] However, due to the unique coherent imaging mechanism of synthetic aperture radar (SAR), the resulting speckle noise causes large fluctuations in pixel values for similar targets. Furthermore, the angular sensitivity of SAR images causes target scattering characteristics to change drastically with azimuth, and terrain interaction induces distortions such as perspective contraction, further amplifying the differences between similar targets while simultaneously making targets of different categories exhibit similar characteristics. The large intra-class differences and high inter-class similarity of targets in SAR images pose a significant challenge to SAR image target recognition.
[0011] Center loss is a loss function used in deep learning to enhance the model's discriminative ability, especially in classification tasks. Its core idea is to improve intra-class compactness and inter-class separation by bringing features of samples from the same class as close as possible to their class center.
[0012] The core idea of center loss is to minimize the distance between sample features and class centers, thereby making samples of the same class more clustered in the feature space. Specifically, for each sample, the Euclidean distance between its feature vector and the corresponding class center is calculated, and a masking operation is used to retain only the distance values of the correct class. Finally, the loss value is calculated.
[0013] The primary function of center loss is to reduce intra-class sample dissimilarity, making samples of the same class more compact in the feature space, thereby improving intra-class similarity. However, it does not directly increase inter-class dissimilarity, so it is usually used in conjunction with softmax loss to achieve both intra-class compactness and inter-class separability. By adjusting the weights of softmax loss and center loss, their contributions can be balanced to improve classification performance.
[0014] Center loss is widely used in deep learning tasks such as face recognition and image classification. Its implementation is typically based on tensor operations and matrix manipulations, with frameworks like PyTorch and TensorFlow providing corresponding implementations. During training, the class centers are dynamically adjusted as samples are updated to ensure continuous model optimization.
[0015] For SAR image target recognition tasks, although the center loss function takes into account the problem of large intra-class differences, it is not very adaptable when dealing with scenarios with high inter-class similarity. Summary of the Invention
[0016] The technical problem to be solved by the present invention is to address the deficiencies involved in the background technology by providing a SAR image target recognition method based on the interval-enhanced center loss function.
[0017] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The SAR image target recognition method based on the interval-enhanced central loss function includes the following steps: Step 1), input the target features and category of the SAR image; Step 2), initialize the center point of each category; Step 3), calculate the Euclidean distance between each target feature and its class center, and average the distances to obtain the intra-class loss; Step 4), calculate the offset of each target feature relative to the class center to update the class center; Step 5) Calculate the Euclidean distance of each target feature to other class centers, find the nearest non-class center distance, construct the maximum margin loss using margin, calculate the positive loss between them, and take the average to obtain the inter-class loss; Step 6) Weighted sum of intra-class and inter-class losses to return the final total loss, forming the interval-enhanced central loss function; Step 7) Jointly train the interval-enhanced center loss function and the cross-entropy loss function to obtain the SAR image target recognition model; Step 8) Input the SAR image to be identified into the SAR image target recognition model for recognition.
[0018] As a further optimization of the SAR image target recognition method based on the interval-enhanced center loss function of the present invention, the training rounds are set to 100 times during joint training in step 7).
[0019] As a further optimization of the SAR image target recognition method based on the interval-enhanced center loss function of the present invention, the SAR image target recognition model is based on the YOLOv8 framework.
[0020] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: This invention introduces inter-class loss on top of the center loss function, forcing a sufficiently large distance between different classes, thus enhancing the model's ability to distinguish different target categories in SAR images. This invention uses a gap-enhanced center loss function combined with a cross-entropy loss function to train the SAR image target recognition network. Experimental results show that on the BistaticMiniSAR dataset, the SAR image target recognition accuracy obtained by this invention can reach 97.59%, which is 0.57% higher than the training result using the center loss function combined with the cross-entropy loss function, and 1.03% higher than the training result using only the cross-entropy loss function. Attached Figure Description
[0021] Figure 1 This is a partial image display of the Bistatic MiniSAR dataset; Figure 2 This is a schematic diagram of the interval-enhanced central loss function. Detailed Implementation
[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: This invention can be implemented in many different forms and should not be considered limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully express the scope of the invention to those skilled in the art. In the drawings, components are enlarged for clarity.
[0023] It should be understood that although the terms first, second, third, etc., may be used herein to describe various elements, components, and / or parts, these elements, components, and / or parts are not limited by these terms. These terms are merely used to distinguish elements, components, and / or parts from one another. Therefore, the first element, component, and / or part discussed below may be a second element, component, or part without departing from the teachings of this invention.
[0024] like Figure 1 As shown, this invention discloses a SAR image target recognition method based on the interval-enhanced center loss function, comprising the following steps: Step 1), input the target features and category of the SAR image; Step 2), initialize the center point of each category; Step 3), calculate the Euclidean distance between each target feature and its class center, and average the distances to obtain the intra-class loss; Step 4), calculate the offset of each target feature relative to the class center to update the class center; Step 5) Calculate the Euclidean distance of each target feature to other class centers, find the nearest non-class center distance, construct the maximum margin loss using margin, calculate the positive loss between them, and take the average to obtain the inter-class loss; Step 6) Weighted sum of the intra-class loss and inter-class loss to return the final total loss, forming the interval-enhanced central loss function, such as... Figure 2 As shown; Step 7) Jointly train the interval-enhanced center loss function and the cross-entropy loss function to obtain the SAR image target recognition model; Step 8) Input the SAR image to be identified into the SAR image target recognition model for recognition.
[0025] When performing joint training in step 7), the number of training rounds is set to 100.
[0026] The SAR image target recognition model is based on the YOLOv8 framework.
[0027] YOLOv8 has made significant progress in the field of object detection. It extends upon previous models such as YOLOv5 and introduces the following novel improvements: 1. Architecture Optimization: YOLOv8 further refines the object detection architecture, improving model efficiency and accuracy. By integrating the improved CSPDarknet backbone and PANet++ neck architecture, better feature extraction and aggregation are achieved. These modifications address gradient redundancy issues and optimize the feature pyramid network, making the model more concise and efficient.
[0028] 2. Model Flexibility: Similar to YOLOv5, YOLOv8 offers multiple model sizes (nano, small, medium, large, extra-large) to adapt to different hardware capabilities and application requirements. The smallest variant, YOLOv8n, is particularly suitable for deployment on edge devices and IoT platforms, providing powerful object detection capabilities with minimal computational overhead.
[0029] 3. Innovative Training Methods: YOLOv8 improves its training methods by introducing advanced data augmentation techniques, such as enhanced mosaic augmentation and adaptive anchor boxes, thereby increasing its ability to detect small objects and reducing its reliance on large datasets. The model also employs mixed-precision training with 16-bit floating-point precision, which accelerates training speed and reduces memory consumption.
[0030] 4. Performance and Impact: YOLOv8 achieves a higher mAP score while maintaining a lower inference time, making it a stronger contender in real-time object detection tasks. Continued use of PyTorch ensures easy access from a broad R&D community, fostering innovation and collaboration in the fields of artificial intelligence and computer vision.
[0031] Cross-entropy loss function is suitable for multi-class classification tasks. In this invention, for the ten-class problem of SAR image target recognition, the interval-enhanced center loss function and cross-entropy loss function are jointly trained, and the two are weighted to obtain the final loss value for each epoch.
[0032] This invention conducted comparative experiments, training using the cross-entropy loss function and training using a combination of the center loss function and the cross-entropy loss function, with epochs set to 100. On the Bistatic MiniSAR dataset, the target recognition accuracy of the SAR image obtained by this method can reach 97.59%, which is 0.57% higher than the training result using the center loss function combined with the cross-entropy loss function, and 1.03% higher than the training result using only the cross-entropy loss function.
[0033] This invention introduces inter-class loss on the basis of the center loss function, which forces a sufficiently large distance between different classes, thereby enhancing the model's ability to distinguish between different target classes in SAR images.
[0034] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.
[0035] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A SAR image target recognition method based on interval-enhanced center loss function, characterized in that, Includes the following steps: Step 1), input the target features and category of the SAR image; Step 2), initialize the center point of each category; Step 3), calculate the Euclidean distance between each target feature and its class center, and average the distances to obtain the intra-class loss; Step 4), calculate the offset of each target feature relative to the class center to update the class center; Step 5) Calculate the Euclidean distance of each target feature to other class centers, find the nearest non-class center distance, construct the maximum margin loss using margin, calculate the positive loss between them, and take the average to obtain the inter-class loss; Step 6) Weighted sum of intra-class and inter-class losses to return the final total loss, forming the interval-enhanced central loss function; Step 7) Jointly train the interval-enhanced center loss function and the cross-entropy loss function to obtain the SAR image target recognition model; Step 8) Input the SAR image to be identified into the SAR image target recognition model for recognition.
2. The SAR image target recognition method based on the interval-enhanced center loss function according to claim 1, characterized in that, When performing joint training in step 7), the number of training rounds is set to 100.
3. The SAR image target recognition method based on the interval-enhanced center loss function according to claim 1, characterized in that, The SAR image target recognition model is based on the YOLOv8 framework.