CN-118736411-B - SAR image arbitrary orientation ship target detection and identification method

Abstract

The invention discloses a detection and identification method for a ship target with an SAR image oriented randomly. The invention adopts a separate regression branch to predict the angle, decouples the angle prediction branch from other branches, realizes more accurate and flexible estimation of the target angle, adopts an integral form representation method proposed in the DFL to predict the angle and the rectangular frame, effectively improves the accuracy of prediction, strengthens the sensitivity of a model to the change of the target size and the direction, further improves the performance in a target detection task, and simultaneously can directly minimize the angle difference between a prediction frame and a real frame in the design of a loss function and remarkably improves the convergence speed of the model when processing the rotation target detection. The invention not only improves the accuracy of the model for detecting the rotating target, but also enhances the adaptability of the model to angle change, thereby providing more robust detection performance when facing the targets with complex background and various postures.

Inventors

• XIA YUANQING
• PAN ZHENHUA
• GAO HAN
• Tang Junlan
• ZHAI DIHUA
• SUN ZHONGQI
• CUI BING
• ZHANG YUAN
• ZHAN YUFENG
• GUO ZEHUA
• DAI LI

Assignees

• 北京理工大学

Dates

Publication Date

20260508

Application Date

20240628

Claims (6)

1. 1. The SAR image arbitrary orientation ship target detection and identification method is characterized by comprising the following steps of: Step 1, constructing a single-stage anchor-free frame network, wherein the Head part of the network comprises an angle regression branch, a rectangular frame regression branch and a classification branch, wherein the angle regression branch is used for predicting the rotation angle of a target, the rectangular frame regression branch is used for predicting the position of the target, and the classification branch is used for predicting the probability that the target belongs to each category; Step 2, training the network constructed in the step 1 based on training set data, wherein the loss function comprises classification loss and regression loss; in the classification loss, the real class of the target is replaced by the predicted frame and the rotation IoU fraction of the real frame in the positive sample, and is 0 in the negative sample; the regression loss comprises a DFL regressing by angles and a rectangular frame, and a rotation frame regression loss riou ;loss riou =1-RIoU, wherein RIoU is rotation-to-transformation ratio; wherein IoU r is the overlapping rate between the predicted frame and the real frame, b and b * are the center points of the predicted frame and the real frame, ρ 2 (b,b * ) is the Euclidean distance between the two center points, ω and ω * are the radian predicted value and the real value of the rectangular frame, w and w * are the widths of the predicted frame and the real frame, and h * are the heights of the predicted frame and the real frame; and step3, detecting and identifying the SAR image ship targets based on the model trained in the step 2.
2. 2. The method of claim 1, wherein the training set data is augmented with a data augmentation policy.
3. 3. The method of claim 2, wherein the data enhancement policy is specifically: And under the condition of not intersecting all other targets in the image, randomly selecting the center point position of the newly added target, keeping the width and the height unchanged, and rotating the target by any angle under the condition that the angle does not exceed the constraint range.
4. 4. The method of claim 3, wherein the rectangular frame is defined by long sides, and wherein the rotation angle of the target is increased in the data enhancement strategy The constraint range of (2) is: Where θ 1 is the angle of the replication target, θ 1 ε [ -pi/2, pi/2).
5. 5. The method of any one of claims 1-4, wherein in step 1, the single-stage anchor-free-frame network is YOLOx, YOLOv6, YOLOv, YOLOv, or FCOS.
6. 6. The method of claim 1, wherein the classification penalty employs a binary cross entropy penalty.

Description

SAR image arbitrary orientation ship target detection and identification method Technical Field The invention relates to the technical field of remote sensing image target detection, in particular to a detection and identification method for a ship target with an arbitrary orientation of SAR images. Background Synthetic aperture radar (SYNTHETIC APERTURE RADAR, SAR) is an active sensing system for active microwave imaging, can penetrate soil and vegetation, is insensitive to cloud cover, and has unique advantages of all-day time, all weather, long acting distance, high resolution and the like. Due to the unique microwave penetrability and high resolution, the target can be reliably detected under severe weather conditions, and the SAR has wide application prospects in the military and civil fields. The ship target detection is taken as an important branch of SAR image interpretation, and has important significance for ocean traffic, fishery management, battlefield situation assessment and military target reconnaissance. According to different methods of SAR image target detection and recognition technology, the method can be roughly divided into two categories of traditional SAR image target detection and SAR image target detection based on deep learning. The traditional SAR image target detection method is mainly based on characteristics and a classifier, and generally consists of three stages of detection, identification and recognition. The traditional detection method requires an expert to rely on own knowledge and experience, and significant features of the image are extracted through manual design, so that the detection step is very complex and cumbersome. At present, deep learning (DEEP LEARNING, DL) shows great advantages in various tasks such as target classification, detection, segmentation and the like by virtue of strong characterization capability and characteristic learning capability. The convolutional neural network (Convolutional Neural Network, CNN) rapidly rises in the field of deep learning due to the strong feature extraction capability, and provides a possible new approach for SAR image target detection research. The current CNN-based target detection method has end-to-end training and reasoning functions, and can be generally divided into two stages and a single stage. The two-stage detector needs to classify the candidate frame and return the boundary frame after obtaining a large number of candidate regions. The single-stage-based target detection network can directly predict the category and the bounding box position information of the target only by looking at the picture once, thereby improving the target detection speed. SAR images are mostly acquired by airborne or spaceborne radars, the imaging angle is generally much larger than that of optical images, and a large imaging span results in some targets being very small in size. SAR images are usually in a bird's eye view angle, and the scene is complex and not intuitive, so that ship targets are usually in any direction and densely arranged near the shore. The horizontal bounding boxes can introduce excessive background in SAR ship target detection, and can also cause adjacent bounding boxes to cover each other, so that missed detection is caused. However, the existing SAR image rotation target detection and identification method does not solve the problems that small-size targets are not easy to detect and the number of targets is small, and the rotation angle is predicted by a classification method, so that the prediction is not fine enough, and the problem of boundary effect is caused. Disclosure of Invention In view of the above, the invention provides a detection and identification method for the ship targets with arbitrary orientation of SAR images, and the detection accuracy and efficiency of small rotating targets in a densely berthing scene of a port ship are effectively realized by simultaneously carrying out regression prediction on the positions and orientations of the ship targets. The SAR image arbitrary orientation ship target detection and identification method comprises the following steps: Step 1, constructing a single-stage anchor-free frame network, wherein the Head part of the network comprises an angle regression branch, a rectangular frame regression branch and a classification branch, wherein the angle regression branch is used for predicting the rotation angle of a target, the rectangular frame regression branch is used for predicting the position of the target, and the classification branch is used for predicting the probability that the target belongs to each category; Step 2, training the network constructed in the step 1 based on training set data, wherein the loss function comprises classification loss and regression loss; in the classification loss, the real class of the target is replaced by the predicted frame and the rotation IoU fraction of the real frame in the positive sample, and is 0 in th