Multi-direction ship target detection and identification method in high-resolution remote sensing image

Abstract

The invention provides a multi-direction ship target detection and identification method in a high-resolution remote sensing image. The method is characterized by designing and realizing a multi-direction ship target detection and identification network structure under a complex scene and in a large-scale environment; automatically exploring shallow-layer and deep-layer multidimensional features and the like of a ship target based on a convolutional neural network; and realizing quick and accurate detection and recognition of multi-directional, and different scales and kinds of ship targets inthe high-resolution remote sensing image under a small sample environment. Compared with a conventional target detection method, the method is faster in detection speed and more accurate in detectionresult.

Description

technical field [0001] The invention relates to the technical field of earth observation, in particular to a detection and recognition method for multi-directional ship targets in high-resolution remote sensing images. Background technique [0002] With the continuous development of earth observation technology in recent years, the acquisition of high-resolution and high-precision remote sensing images has become easier. High-resolution remote sensing images provide a good foundation for more accurate target detection and recognition. Ships are important targets at sea, and their detection and identification are of great significance in both civil and military fields. In terms of civilian use, the detection and identification of ships can monitor the position of ships at sea, which is convenient for maritime search and rescue, search of smuggling ships, and monitoring of sea areas; [0003] At present, there is no perfect large-scale ship target detection and recognition d...

AI Technical Summary

Problems solved by technology

[0007] For the detection and recognition of high-resolution remote sensing ship targets, most of the existing image sample generation and augmentation methods are based on small scene images (resolution below 1920x1080), and there is no sample generation method for large scene images. Meet the generation and augmentation requirements of deep learning training samples in remote sensing images of large scenes

[0008] In traditional object detection and recognition methods, the use of expert features needs to describe and design features based on the prior knowledge of experts, which is highly subjective and easily affected by expert experience.

In addition, traditional machine learning classification models cannot make full use of large-scale data information, and it is difficult to achieve satisfactory classification results, resulting in a decrease in the accuracy of target detection and target recognition.

[0009] The existing deep learning methods cannot meet the technical requirements of multi-oriented ship target detection and recognition, mainly in the following aspects: first, the detection results of these methods use horizontal rectangular boxes to describe the position of the detection target, which is not suitable for multi-oriented ship targets. The description of the target position is not accurate enough, and it will contain more background information for non-horizontal and vertical ship targets

Second, the feature extraction layer used in these methods is relatively simple, and the aspect ratio and scale settings of the prior frame are relatively simple

Therefore, the existing neural network algorithms cannot detect large-scale, multi-aspect ratio aircraft carriers, ship targets or small-scale targets

Third, some methods (such as DRBox) detect multi-oriented targets in remote sensing images by designing multi-oriented prior boxes, but they use a unique feature map as the feature extraction layer and do not make full use of features of different dimensions

In order to detect targets of different scales, the image pyramid is established by scaling the image at different scales, and input to the network to detect targets of different scales in the same image. The computational complexity is high, the efficiency is low, and the time is long.

Fourth, these methods are limited by the size of the input image and cannot directly detect targets in remote sensing images of large scenes

The size of the input image of the deep learning network is fixed, while the remote sensing image of a large scene has a large size and high resolution. If the image is directly scaled, the ship target will be scaled to a point, which will cause the detection algorithm to fail

Fifth, only a single target is detected, and it is useless to make full use of the classification information of different targets

In the training process, the parameters of the model are not corrected by the loss of the category, and it does not have the ability to identify different categories of targets.

Images