A remote sensing image detection method and system

Abstract

A remote sensing image detection method and system, input picture, after backbone network, feature pyramid network (FPN), region proposal network (RPN) and region of interest alignment (RoIAlign) network, spatial information and position information are obtained; the spatial information is encoded from the dimension of the channel, and the spatial and channel features are reorganized; the position information is transmitted to the classification converter to obtain the position information converter, the significant information in the classification target is discriminated, and the global modeling class feature is obtained; the regression branch detection head is constructed based on the spatial and channel feature reorganization, the classification branch detection head is constructed based on the position information converter, and image detection is carried out in parallel. In the positioning branch of the detection head, the spatial channel information feature reorganization is carried out, the spatial information and the channel information are modeled to solve the feature annihilation and loss problem in the feature extraction process in the positioning process. In the classification branch, the category detection head based on the Transformer is proposed, and the feature difference of different targets is captured from the global information.

Images