A Ship Detection Method in Satellite Imagery Combining Rotating Frame and Context Information

Abstract

The invention provides a satellite image ship detection method combined with rotating frame and context information. Firstly, color transformation, scale transformation and angle transformation are used to expand the training data set; Semantic information and feature maps with precise location information; then use the rotating region candidate network to generate a series of multi-scale, multi-aspect ratio, and multi-angle candidate bounding boxes on the feature map; finally use the multi-region fully connected network to extract candidate bounding boxes Context composite features are used for ship detection. Compared with the previous ship detection methods, the present invention can obtain higher detection accuracy from multiple perspectives such as data preprocessing, feature extraction, candidate bounding box generation, and detectors.

Description

technical field [0001] The invention belongs to the field of satellite image recognition, and more particularly relates to a satellite image ship detection method combined with rotating frames and context information. Background technique [0002] With the rapid development of remote sensing technology, current satellite sensors can easily acquire high-resolution images. Ship detection using high-resolution satellite images is an important research direction, and has been widely used in fishery management, ship traffic services, maritime operations and other fields. However, due to some significant differences between satellite images of ship targets and targets in natural images, such as: (1) the illumination changes, scale changes and aspect ratio changes of satellite image ship targets are more intense; (2) ship targets will encounter a large number of different complex environment, such as near-coastal environment, far-coastal environment, and interference from targets ...

AI Technical Summary

Problems solved by technology

However, these methods are specially designed for object detection in natural images, and it is often difficult to achieve ideal results when directly applied to satellite images.

Images