Remote sensing target detection method based on content awareness

Abstract

The invention provides a remote sensing target detection method based on content awareness. The method comprises the following steps: obtaining coordinate information of a prediction frame according to information output by a regression module, and obtaining coordinate information of a real frame from an original remote sensing image; positioning on the maximum feature map according to the coordinate information of the prediction frame and the coordinate information of the real frame to obtain feature information v* and v contained in the prediction frame and the real frame respectively; determining a common minimum horizontal enclosing rectangle of the prediction frame and the real frame, and taking out an area where the minimum horizontal enclosing rectangle is located from the maximum feature map, and recording the area as a feature map x; respectively setting contents which do not belong to the real frame and the prediction frame on the feature map x to be 0 to obtain a feature map f1 corresponding to the real frame and a feature map f1* corresponding to the prediction frame; calculating the similarity between the feature map f1 and the feature map f1*, taking the similarity as a content consistency loss Lfeed, and constraining the regression of the anchor frame based on the content consistency loss Lfeed.

Description

technical field [0001] The invention relates to the technical field of satellite remote sensing, in particular to a content-aware-based remote sensing target detection method. Background technique [0002] Horizontal object detectors, such as R-CNN, fastR-CNN, faster R-CNN, YOLO, are designed for detecting horizontal objects. These methods usually use horizontal bounding boxes (HBB) to capture objects in natural images. Different from horizontal object detection tasks, arbitrary orientation object detection relies on oriented bounding boxes (OBBs) to capture objects in arbitrary orientations. Current object-oriented detection methods are generally extended from horizontal object detectors. For example, R2CNN uses a Region Proposal Network (RPN) to produce the HBB of text, and incorporates different scales of ensemble ROI features to regress the parameters of the OBB. R2PN incorporates bounding box orientation parameters into the RPN network, and develops a rotated RPN net...

AI Technical Summary

Problems solved by technology

However, on the same target object, multiple suggested regression boxes with the same IoU can be obtained, but the content information contained in these regression boxes is different, and the inventor found that the content information contai...

Images