基于多尺度特征注意力机制的轻量化室内定位模型OA

To address the issues of narrow receptive fields and insufficient global feature learning in existing indoor localization models,this paper proposes a lightweight multi-scale feature fusion indoor localization model,MLGNet.A multi-scale feature pyramid module is designed,which employs multi-branch dilated convolution and asymmetric convolution to expand the model's receptive fields.Meanwhile,a kernel function-enhanced self-attention module and a dynamic aggregation channel attention module are constructed to achieve efficient feature extraction in both global and channel dimensions.Finally,a multi-stage feature fusion strategy is adopted to integrate local and global features step by step.Compared with existing models,MLGNet exhibits more than 15%localization accuracy improvement on the CTW and KU Leuven datasets.The number of model parameters is only 2.58 M.