Indoor camera relocation method based on composite attention guidance

The invention discloses an indoor camera relocation method based on composite attention guidance. The method is specifically implemented according to the following steps: step 1, image feature extraction; step 2, multi-network hierarchical feature fusion; 3, reconstructing a three-dimensional scene code; step 4, performing feature extraction by the backbone network; step 5, semantic feature fusion; step 6, recovering image details; step 7, neural network training optimization; and step 8, model strategy training optimization. According to the method, feature information of different levels is integrated through a multi-level feature fusion module, feature maps of different levels are perfected by embedding an embedded attention module (EAM), space details and channel information of low-level feature maps are enhanced, features are decoupled from sparse textures and repeated scenes through multi-level context fusion, and the feature fusion efficiency is improved. Therefore, stable and accurate visual positioni