To address the challenges of insufficient feature extraction and artifact generation in lightweight depth estimation methods caused by parameter reduction

a self-supervised monocular depth estimation framework incorporating a lightweight guided filtering module(LGFM)

is proposed.The framework is composed of DepthNet and PoseNet

and a dynamic guided filtering(DGF)module is introduced

which incorporates a lightweight dynamic guided filter kernel(DGFK)generation module and edge-guided filtering(EGF).The filtering module uses the Sobel operator to extract edge information from RGB images

and combines it with dynamically generated filter kernels

which enhances the model’s ability to recover details and edge information and effectively suppresses artifacts.Moreover

the number of model parameters is significantly reduced through channel dimension reduction and residual connections.Experimental results show that with 1.52M parameters

the LGFM framework achieves better performance than existing methods on the KITTI dataset

particularly in detail recovery and artifact reduction.