CN-115761527-B - SAR image target detection method based on fusion of statistical characteristics and structural characteristics

Abstract

The invention discloses a SAR image target detection method based on the fusion of statistical characteristics and structural characteristics, which comprises the steps of carrying out target recognition on an SAR image to be recognized by utilizing a CFAR detector and a Faster-RCNN detection network in an alternating iteration mode until an alternating iteration termination condition is met, so as to obtain a target recognition result of the SAR image to be recognized; according to the invention, the CFAR detector and the fast-RCNN detection network are fused, a large number of false alarm boundary boxes are obtained according to the CFAR detector, soft labels of the boundary boxes are obtained through fusion calculation from pixel level to target level, the fast-RCNN detection network is used as a basic training frame, the boundary boxes output by the fast-RCNN detection network are used as the protection area size of the CFAR detector, the mixing of target samples in clutter samples is reduced, and the parameter estimation of the statistical distribution model is more accurate.

Inventors

• Wen Zaidao
• YANG TAO
• LIU ZHUNGA
• PAN QUAN

Assignees

• 西北工业大学

Dates

Publication Date

20260505

Application Date

20221207

Claims (6)

1. 1. The SAR image target detection method based on the fusion of the statistical characteristics and the structural characteristics is characterized by comprising the following steps of: Performing target recognition on the SAR image to be recognized by using the CFAR detector and the fast-RCNN detection network in an alternating iteration mode until an alternating iteration termination condition is met, and obtaining a target recognition result of the SAR image to be recognized; training the fast-RCNN detection network by utilizing a target recognition result output by the CFAR detector, and utilizing a bounding box output by the fast-RCNN detection network as a protection area size of the CFAR detector; Training the fast-RCNN detection network using the target recognition results output by the CFAR detector includes: adding a target identification result output by the CFAR detector as a weight into a two-class cross entropy loss function of a fast-RCNN detection network; performing target recognition on the SAR image to be recognized by using the CFAR detector comprises the following steps: combining the connected target pixels in the SAR image to be identified to obtain a first target frame; The first target frame is enlarged in proportion to obtain a second target frame; The second target frame is obtained, and then the method further comprises the following steps: fusing the confidence degrees of all pixels in the second target frame to obtain the confidence degrees of the second target frame; training the fast-RCNN detection network using the target recognition result output by the CFAR detector: and taking the confidence coefficient of the second target frame as a target recognition result output by the CFAR detector.
2. 2. The SAR image target detection method based on the fusion of statistical properties and structural properties of claim 1, wherein the statistical distribution model selection method of the CFAR detector comprises: calculating the deviation between the SAR image to be identified and the statistical distribution model; And selecting a statistical distribution model corresponding to the deviation minimum value as the statistical distribution model of the CFAR detector.
3. 3. The SAR image target detection method based on the fusion of statistical properties and structural properties of claim 2, wherein the shape parameter and the scale parameter of the statistical distribution model are calculated by using a maximum likelihood estimation method.
4. 4. The SAR image target detection method based on the fusion of statistical properties and structural properties according to claim 3, wherein the loss function of the fast-RCNN detection network is: , Wherein, the For the purpose of the RPN loss, Head loss was detected for Faster-RCNN.
5. 5. The SAR image target detection method based on the fusion of statistical properties and structural properties of claim 4, wherein the RPN loss is: , Wherein, the For the purpose of the two-class cross entropy loss, , , In an infinitesimal amount of the solution, The confidence common label for the jth target, The confidence predictive value of the RPN to the jth target object is given, N is the number of the target objects, For the weight corresponding to each tag, , For the RPN target frame loss, For the size of the second target frame, Is a predicted value of the RPN candidate bounding box size.
6. 6. The SAR image target detection method based on the fusion of statistical properties and structural properties according to claim 5, wherein the fast-RCNN detection head loss is: , Wherein, the KL divergence for the fast-RCNN detection head, In order to detect the head-to-class probability prediction output, In order to detect the loss of the head target frame, Is a predicted value for predicting the bounding box size.

Description

SAR image target detection method based on fusion of statistical characteristics and structural characteristics Technical Field The invention belongs to the technical field of SAR image target detection, and particularly relates to a SAR image target detection method based on fusion of statistical characteristics and structural characteristics. Background The reliable ocean monitoring capability can effectively ensure the ocean rights of a country, and can be beneficial to the development of tasks such as offshore rescue, fishery management, offshore traffic control and the like. When monitoring and detecting the ship targets in the unfamiliar sea area by using the space-borne SAR, the marked ship targets lack corresponding expert knowledge (such as ship GPS information) support. Because the space-borne SAR has wide coverage area, the position of the target object is sparsely distributed. The manual vision is only used for marking the ship data, so that time and labor are wasted, moreover, the artificial judgment of the target object is easily interfered by sea clutter and other factors, and the marking of the ship target is very difficult. Therefore, the detection of the target of the spaceborne SAR ship is often faced with the condition of unsupervised information. With the development of artificial intelligence technology, deep learning is widely applied to image target detection. However, fully supervised target detection algorithms based on deep learning (e.g., faster-RCNN, YOLO-V3, etc.) require a large number of accurate instance labels. Most current unsupervised target detection is based on mining the differences between the previous and subsequent frames and frame images from the video to obtain target location information. While on-board SAR is typically dominated by single-shot image detection, such methods are not applicable. The detector based on deep learning has high detection speed and can extract and utilize the structural features of the image, but a large number of instance level labels are needed. Conventional SAR-characteristic-based methods typically do not require labeling information. For example, the method based on SAR image polarization features is mainly based on the fact that ship targets and ocean scattering mechanisms are different, and therefore the targets and the background can be distinguished by the polarization features. However, for unipolar data, the method described above is not applicable. The method based on the statistical characteristics of SAR plays an important role in ship detection, and the algorithm is mainly a constant false alarm detection CFAR detection algorithm and a series of methods derived from the constant false alarm detection CFAR detection algorithm, such as bilateral CFAR (TP-CFAR), mean value CFAR (CA-CFAR) and the like. The method comprises the steps of firstly establishing a mathematical statistical model for clutter power, then solving a threshold T for accepting or rejecting the hypothesis under the condition of setting a certain false alarm rate by a hypothesis test method, and judging whether a pixel point is a target object or not. However, the method based on constant false detection needs to traverse the pixel points by using a detection unit, has low detection speed and is easy to generate false alarm. Disclosure of Invention The invention aims to provide a SAR image target detection method based on the fusion of statistical characteristics and structural characteristics, so as to reduce CFAR false alarm rate and improve target detection precision. The SAR image target detection method based on the fusion of the statistical characteristics and the structural characteristics comprises the following steps: Performing target recognition on the SAR image to be recognized by using the CFAR detector and the fast-RCNN detection network in an alternating iteration mode until the alternating iteration termination condition is met, and obtaining a target recognition result of the SAR image to be recognized; The method comprises the steps of training a Faster-RCNN to detect a network by using a target identification result output by a CFAR detector, and updating the protection area size of the CFAR detector by using a target identification result output by the Faster-RCNN to detect the network. Further, the statistical distribution model selection method of the CFAR detector comprises the following steps: calculating the deviation between the SAR image to be identified and the statistical distribution model; And selecting a statistical distribution model corresponding to the minimum deviation value as the statistical distribution model of the CFAR detector. Further, the shape parameters and the scale parameters of the statistical distribution model are calculated by using a maximum likelihood estimation method. Further, training the fast-RCNN to detect the network using the target recognition result output by the CFAR detector includes: and addin