CN-121982556-A - Deep learning-based optical satellite image ship detection and course judgment method

CN121982556ACN 121982556 ACN121982556 ACN 121982556ACN-121982556-A

Abstract

The application relates to the technical field of image recognition and computer vision, and discloses a method for detecting and judging the course of a ship by using an optical satellite image based on deep learning, which comprises the steps of obtaining a high-resolution optical satellite image to be processed; the method comprises the steps of carrying out sliding window segmentation on a high-resolution optical satellite image to obtain a plurality of image blocks, sequentially inputting the image blocks into a pre-trained end-to-end detection network to obtain ship target prediction results of each image block, wherein the prediction results comprise position information and heading information, the end-to-end detection network comprises a backbone network, a neck network and a decoupling detection head which are sequentially connected, decoding and overall de-duplication processing are carried out on all the obtained prediction results, and final ship detection and heading judgment results are generated. The application solves the technical problems of accurate detection and reliable heading judgment of the ship targets in the high-resolution optical satellite images through the end-to-end deep learning architecture, and realizes remarkable improvement of precision, efficiency and complete functions.

Inventors

HUANG GUANXIAN
Qilin Spring Batch
SUN SHAOJIE
XIAO HONG
QU ZIHONG
LIN XIAOBO
ZHAO JUN
WANG JIALIN
Xu Gengran
SHI XIAOCHUN

Assignees

广东省国土资源测绘院
中山大学

Dates

Publication Date: 20260505
Application Date: 20251209

Claims (10)

1. The method for detecting and judging the course of the optical satellite image ship based on the deep learning is characterized by comprising the following steps: Acquiring a high-resolution optical satellite image to be processed; Carrying out sliding window segmentation on the high-resolution optical satellite image to obtain a plurality of image blocks; The method comprises the steps of inputting image blocks into a pre-trained end-to-end detection network in sequence to obtain a ship target prediction result of each image block, wherein the prediction result comprises position information and heading information, the end-to-end detection network comprises a backbone network, a neck network and a decoupling detection head which are sequentially connected, and the decoupling detection head comprises at least two independent branches, namely a position detection branch for detecting a target position and a heading estimation branch for estimating the heading; and decoding and global de-duplication processing are carried out on all the obtained prediction results, and final ship detection and course judgment results are generated.
2. The method for detecting and judging the ship by the optical satellite image based on the deep learning is characterized in that the course estimation branch adopts a discrete classification and continuous regression joint representation method, and specifically comprises a discrete classification sub-branch and a continuous regression sub-branch, wherein the discrete classification sub-branch is used for discretizing a 360-degree range of a course angle into N predefined categories and outputting probability distribution of the category to which the ship course belongs, the continuous regression sub-branch is used for predicting continuous offset relative to a corresponding angle interval in a discrete classification result, and the final course angle of the ship is obtained by calculating the offset of the angle interval output by the discrete classification sub-branch and the continuous regression sub-branch.
3. The method for detecting and judging the ship course by using the optical satellite images based on the deep learning according to claim 2, wherein the value of N is 36, the 360-degree course angle is divided into 36 continuous categories, and the angle span of each category is 10 degrees.
4. The deep learning-based optical satellite image ship detection and course judgment method is characterized in that a light-weight network structure constructed based on a C3k2 module is adopted by a backbone network, the backbone network is used for extracting a multi-scale feature map from an input image block, and the multi-scale feature map comprises a shallow feature map with high spatial resolution and a deep feature map with strong semantic information.
5. The deep learning-based optical satellite image ship detection and course judgment method according to claim 4 is characterized in that a neck network adopts a structure based on a path aggregation network PANet and is combined with a space pyramid rapid pooling SPPF module and a channel space attention C2PSA module, and the neck network is used for carrying out bidirectional fusion on the multi-scale feature images from the backbone network to generate a multi-scale fusion feature image with high semantic information and high spatial resolution.
6. The method for detecting and judging the ship by using the optical satellite image based on the deep learning according to claim 1, wherein the decoupling detection head further comprises a target classification branch for predicting whether a ship target and a ship type belong to the ship target exist in an image, and the target classification branch, the position detection branch and the heading estimation branch share the characteristics extracted by the backbone network and the neck network during training, but perform task-specific characteristic learning and prediction through respective independent convolution layers.
7. The method for detecting and judging the heading of the optical satellite image ship based on deep learning according to claim 3, wherein the end-to-end detection network performs end-to-end training through a multi-task loss function, the multi-task loss function at least comprises target classification loss, position regression loss and heading loss, and the heading loss is a weighted sum of classification loss of the discrete classification sub-branch and regression loss of the continuous regression sub-branch.
8. The method for detecting and judging the course of the optical satellite image ship based on the deep learning according to claim 1, wherein the sliding window segmentation adopts windows with overlapping areas for sampling so as to prevent the ship target from being cut by the edges of the windows.
9. The method for detecting and determining the heading of a ship based on deep learning of any one of claims 1 to 8, wherein the global de-duplication process uses a global non-maximum suppression algorithm to eliminate redundant frames generated by repeated detection of the same ship target by different sliding windows.
10. An optical satellite image ship detection and heading judgment system based on deep learning, which is used for realizing the method as claimed in any one of claims 1 to 9, and comprises the following steps: the image acquisition module is configured to acquire a high-resolution optical satellite image to be processed; the preprocessing module is configured to perform sliding window segmentation on the high-resolution optical satellite image to obtain a plurality of image blocks; The system comprises a detection module, a ship target prediction result acquisition module and a ship target prediction module, wherein the image blocks are sequentially input into a pre-trained end-to-end detection network to obtain a ship target prediction result of each image block, the prediction result comprises position information and course information, the end-to-end detection network comprises a backbone network, a neck network and a decoupling detection head which are sequentially connected, and the decoupling detection head comprises at least two independent branches, namely a position detection branch for detecting a target position and a course estimation branch for estimating a course; And the processing module is configured to decode and globally de-duplicate all the obtained prediction results to generate final ship detection and course judgment results.

Description

Deep learning-based optical satellite image ship detection and course judgment method Technical Field The application relates to the technical field of image recognition and computer vision, in particular to a method for detecting and judging heading of an optical satellite image ship based on deep learning. Background With the rapid development of high-resolution optical satellite remote sensing technology, the ship target detection based on satellite images has an important application value in the fields of marine traffic monitoring, fishery management, military reconnaissance and the like. However, the prior art still faces a series of technical challenges in practical applications. Early ship detection methods relied primarily on traditional image processing techniques and classical machine learning algorithms. The first type of method is a threshold segmentation-based technique that uses the difference in gray scale or texture between a ship and a sea background for segmentation. The method is extremely sensitive to environmental interference, and is easy to generate a large number of false alarms and missed detection under the conditions of illumination change, cloud cover, wave fluctuation or solar flare and the like, so that the robustness is poor. The second type of method is a machine learning method based on hand design features, for example, features such as a directional gradient histogram, scale-invariant feature transformation or a local binary pattern are extracted by adopting a sliding window, and are distinguished by combining with a classifier such as a support vector machine or adaptive enhancement. Although the performance of the method is improved compared with threshold segmentation, the design of manual characteristics is highly dependent on expert priori knowledge, the characterization capability is limited, and the method is difficult to fully cope with the complex and changeable marine environment and the diversity of the ship self-morphology. Meanwhile, the calculation paradigm of sliding window traversal is low in efficiency, and is difficult to meet the rapid processing requirements of large-breadth and high-resolution satellite images. In recent years, target detection algorithms based on deep convolutional neural networks, such as fast R-CNN series based on candidate regions and YOLO and SSD series based on single detection, have become the mainstream technology in the field by virtue of their strong automatic feature learning capability. However, the direct application of such a general object detection model to the ship detection task of high-resolution optical satellite images still has the limitation that dense small objects are difficult to detect, and the pixel area occupied by the ship objects (especially small fishing vessels) in the high-resolution images is usually very small, which belongs to typical small objects. The existing general detection model needs to be downsampled for multiple times in the feature extraction process, so that shallow features have abundant space details but insufficient semantic information, and deep features have abundant semantic information but lose a large amount of fine space structures. This contradiction is extremely prone to missing inspection of small-sized ships densely distributed in the image. The key dynamic information perception is lacking, and most of the existing methods can only output a horizontal rectangular boundary box representing the ship position, and lack the perception capability of the key dynamic attribute of the ship course (i.e. the movement direction). Heading information is critical to judging ship intent and distinguishing behavior modes (such as sailing, berthing and working). While few studies have attempted to introduce a rotating bounding box to characterize direction, it is limited by model structural design, which tends to be less accurate in predicting heading angle. Especially under the condition that the morphological characteristics of the heads and the tails of the ships are similar, the model is extremely easy to generate 180-degree reverse misjudgment, the reliability is insufficient, and the requirements of the application such as offshore traffic control and illegal behavior supervision on dynamic information analysis are difficult to meet. Complex background interference and model efficiency bottleneck, namely that a large number of interferents similar to ship characteristics exist in the ocean background, such as broken cloud layers, sea wave white crowns, island submerged reefs and the like, so that the model false alarm rate is high. On the other hand, the high-resolution satellite image data size is huge, and the complex depth model designed for pursuing high precision is huge in calculation cost, so that the processing speed is low, the application requirements on quick and near-real-time processing of remote sensing data are difficult to meet, and the difficulty and cost for deploy