A frequency domain enhanced self-adaptive visual text alignment remote sensing weakly supervised segmentation method

Abstract

The present application relates to a kind of frequency domain enhancement adaptive visual text alignment remote sensing weak supervision segmentation method, the method comprises the following steps: obtaining remote sensing image and category text data;Frequency domain enhancement adaptive visual text alignment remote sensing weak supervision segmentation model is built;Frequency domain enhancement adaptive visual text alignment remote sensing weak supervision segmentation model is trained;Obtain the remote sensing image to be segmented and category text;Obtain pixel-level semantic segmentation result.Compared with prior art, the present application is decomposed by frequency domain enhancement module to remote sensing image adaptive high-frequency and low-frequency component, while retaining the edge details of ground object target, suppresses noise interference, effectively alleviates the boundary fuzzy problem under weak supervision condition, realizes the text representation of image adaptation by multi-dimensional characteristic fusion, avoids the disconnection between category text and image content, and simultaneously uses bidirectional alignment to realize the mutual calibration of text and vision, effectively improves the positioning accuracy and boundary integrity of remote sensing semantic segmentation.

Images