CN114219996A

CN114219996A - SAR image berthing ship target detection method

Info

Publication number: CN114219996A
Application number: CN202111606610.2A
Authority: CN
Inventors: 谢聪; 庄龙; 郭杰
Original assignee: CETC 14 Research Institute
Current assignee: CETC 14 Research Institute
Priority date: 2021-12-26
Filing date: 2021-12-26
Publication date: 2022-03-22

Abstract

At present, the mainstream SAR ship detection method still has many unsolved difficult problems, for example, on land and artificial buildings near the coastline, such as a gallery road, a house, a tower and the like, strong scatterers cause great interference to ship detection, and part of regions have scattering characteristics similar to those of ships, so that the existing method has poor ship detection effect near the coastline, has high false alarm rate, and is difficult to accurately detect the position and the contour of the offshore ships. The invention provides a detection method for a ship target berthed by an SAR image, which not only can accurately obtain the position, the direction, the outline and the characteristic information of a ship berthed near a coastline, but also can reduce the influence of strong scatterers such as land and artificial buildings near the coastline on the detection of the ship.

Description

SAR image berthing ship target detection method

Technical Field

The invention belongs to the field of radar target detection, and particularly relates to a detection method for a target berthed on a ship by an SAR image.

Background

A method for rapidly detecting a ship target based on a large-scene remote sensing image of a cascaded convolutional neural network (radar science and newspaper) discloses a cascaded convolutional neural network detection framework, wherein the detection framework is formed by cascading two full convolutional networks, namely a target pre-screening full convolutional network (P-FCN) and a target accurate detection full convolutional network (D-FCN). The P-FCN is a lightweight image classification network and is responsible for quickly pre-screening possible ship areas in a large scene image, the number of layers is small, training is simple, redundancy of candidate frames is small, and calculation burden of a subsequent network can be reduced; the D-FCN is an improved U-Net network, and the target mask and the ship orientation estimation layer are added into a traditional U-Net structure to perform multi-task learning, so that the fine positioning of any ship orientation target is realized. An SAR ship target detection method based on a deep convolutional neural network becomes a mainstream method in the field. "Joint connected network for small-scale ship classification in SAR images" ("IGARSS") discloses a convolutional neural network combining a generator network and a classification network, wherein the generator network is used for learning a mapping function for converting a low-resolution ship image into a high-resolution ship image, and the classification network classifies the generated high-resolution ship image as input. A joint loss optimization strategy is provided for training and optimizing the model, and experiments show that the ship picture generated by the generator effectively improves the classification precision of the model. At present, the mainstream SAR ship detection method still has many unsolved difficult problems, for example, on land and artificial buildings near the coastline, such as a gallery road, a house, a tower and the like, strong scatterers cause great interference to ship detection, and part of regions have scattering characteristics similar to those of ships, so that the existing method has poor ship detection effect near the coastline, has high false alarm rate, and is difficult to accurately detect the position and the contour of the offshore ships.

Disclosure of Invention

In order to overcome the defects in the prior art, the invention provides the SAR image berthing ship target detection method, which not only can accurately obtain the position, the direction, the contour and the characteristic information of a berthing ship near a coastline, but also can reduce the influence of strong scatterers such as land, artificial buildings and the like near the coastline on the ship detection. The method comprises the following specific steps:

step (1): image pre-processing

Image enhancement: the brightness and the contrast of the SAR image are enhanced by adopting an image gray level histogram equalization method, so that the detailed information of the image is displayed more obviously;

denoising an image: then, denoising pretreatment is carried out on the SAR image by using an adaptive denoising algorithm based on NSCT, so that speckle noise is removed and image detail texture information is protected.

Step (2): naval vessel target coarse detection

A deep convolutional neural network-based SAR image ship target detection method becomes a mainstream method in the field, however, in the deep neural network, small target characteristic information is easy to lose, and finally small target detection omission is caused. Aiming at the problem, a network structure (SSS-YOLOv 3) for a small target is designed, a cavity convolution module is introduced to expand a perception field, and preliminary detection of the SAR image ship target is realized.

And (3): sea-land segmentation and false alarm rejection

On the basis of ship target coarse detection, sea and land segmentation is carried out on the SAR image by adopting a multi-scale OTSU threshold segmentation method. Firstly, binary segmentation is carried out on the SAR image by utilizing a global OTSU threshold, and connectivity area extraction is carried out on the segmentation result of the OTSU threshold, so that a foreground communication area and a background communication area are obtained. Secondly, performing threshold operation on the area of the foreground communication area, and setting the foreground area smaller than a certain area as a background area to remove the influence of sea clutter.

On the basis of sea and land segmentation, morphological filtering processing is carried out on a sea and land segmentation image, a connected domain analysis method is adopted to segment a ship and land adhesion area, and strong scatterers such as land and artificial buildings near a coastline are eliminated through thresholds such as an aspect ratio and an area.

Step (4) precise detection of ship target based on geometric contour estimation

The minimum circumscribed rectangle of the ship target is extracted by using a geometric contour estimation method, the contour boundary of the ship is obtained, the geometric features (length, width, length-width ratio, area, circumference and the like), the shape features (shape index and the like), the edge features, the scattering features and the like of the ship are respectively calculated by using a multi-layer feature fusion method, and the accurate detection of the ship target is realized.

The invention has the beneficial effects that:

(1) aiming at the problems of poor detection effect, high false alarm rate and the like of a traditional ship detection method for ships berthing near a coastline, a multi-scale threshold segmentation and morphology processing method is adopted, so that strong scatterers such as land, artificial buildings and the like near the coastline can be filtered, and the detection precision of ships berthing near the coastline is improved;

(2) aiming at the problem that the traditional detection method can only obtain a horizontal detection frame of a ship and cannot accurately estimate the position and the direction of a ship target, the minimum circumscribed rectangular frame of the ship target is obtained by using a target contour estimation method, the information such as the contour, the dimension and the like of the ship target can be accurately obtained, and the characteristic information such as the geometric characteristic, the shape characteristic, the edge characteristic, the scattering characteristic and the like of the ship target is extracted by using a multilayer characteristic fusion method, so that the more accurate detection and identification of the offshore berthing ship target are realized.

Drawings

FIG. 1 is a flow chart of the method of the present invention.

Fig. 2 is a schematic diagram of an embodiment.

Detailed Description

The invention is described in further detail below with reference to figures 1 and 2.

(1) Image pre-processing

Selecting three typical satellite-borne SAR images, and improving the brightness and contrast of the SAR images through image gray level histogram equalization; then, speckle noise of the SAR image is eliminated by using an NSCT-based adaptive denoising algorithm, and image edge information can be reserved to a greater extent.

(2) Preliminary detection of ship position

On the basis of SAR image preprocessing, a deep convolutional neural network model (such as SSS-YOLOv 3) is adopted, and a hole convolution module is introduced to expand a perception field of view, so that preliminary detection of SAR image ship targets is realized;

(3) sea-land segmentation and false alarm rejection

On the basis of primary detection of ships, firstly, sea and land segmentation is carried out on SAR images by adopting a multi-scale OTSU threshold segmentation method to obtain ship and land areas; and secondly, performing morphological filtering processing on the sea and land segmentation image, segmenting the adhesion area of the ship and the land by adopting a connected domain analysis method, and rejecting the strong scatterers near the coastline, such as the land and the artificial building, through thresholds such as the length-width ratio, the area and the like.

(4) Precision detection of ship targets

The minimum circumscribed rectangle of the ship target is extracted by using a geometric contour estimation method, the contour boundary of the ship is obtained, the geometric features (length, width, length-width ratio, area, circumference and the like), the shape features (shape index and the like), the edge features, the scattering features and the like of the ship are respectively calculated by using a multi-layer feature fusion method, and the ship target detection is completed.

The present invention is not limited to the above-described specific embodiments, and various modifications and variations are possible. Any modifications, equivalents, improvements and the like made to the above embodiments in accordance with the technical spirit of the present invention should be included in the scope of the present invention.

Claims

1. A SAR image berthing ship target detection method is characterized in that: the method comprises the following steps:

step (1): image pre-processing

denoising an image: then, denoising pretreatment is carried out on the SAR image by using an adaptive denoising algorithm based on NSCT, so that speckle noise is removed and image detail texture information is protected;

step (2): naval vessel target coarse detection

On the basis of SAR image preprocessing, a deep convolution-based neural network model is adopted, and a hole convolution module is introduced to expand a perception field of view, so that preliminary coarse detection on a ship target of an SAR image is realized;

and (3): sea-land segmentation and false alarm rejection

On the basis of ship target coarse detection, sea and land segmentation is carried out on the SAR image by adopting a multi-scale OTSU threshold segmentation method;

on the basis of sea and land segmentation, performing morphological filtering processing on a sea and land segmentation image, segmenting a ship and land adhesion area by adopting a connected domain analysis method, and rejecting strong scatterers such as land and artificial buildings near a coastline through an aspect ratio and an area threshold;

step (4) precise detection of ship target based on geometric contour estimation

The minimum circumscribed rectangle of the ship target is extracted by using a geometric contour estimation method, the contour boundary of the ship is obtained, the geometric characteristics, the shape characteristics, the edge characteristics and the scattering characteristics of the ship are respectively calculated by using a multi-layer characteristic fusion method, and the accurate detection of the ship target berthed is realized.

2. The SAR image-berthing ship target detection method of claim 1, characterized in that: the sea-land segmentation is carried out on the SAR image by adopting a multi-scale OTSU threshold segmentation method, which specifically comprises the following steps:

firstly, binary segmentation is carried out on the SAR image by utilizing a global OTSU threshold, and connectivity area extraction is carried out on the segmentation result of the OTSU threshold to obtain a foreground communication area and a background communication area; secondly, performing threshold operation on the area of the foreground communication area, and setting the foreground area smaller than a certain area as a background area to remove the influence of sea clutter.