Abstract: Reconstructing high-quality meshes from binary voxel data is a long-standing problem in computer graphics. Existing approaches typically rely on feature extraction with fixed receptive fields and geometric priors. However, the low information density and strong discreteness of binary voxels make them ill-suited for directly reconstructing high-quality surfaces, particularly when representing complex geometric struc-tures and capturing long-range edge features. This often leads to stair-step artifacts and the loss of sharp details.  To address this problem, this paper proposes a multi-scale dynamic dual contouring mesh recon-struction method. First, using a multi-scale structure, the mesh is progressively refined, and a residual ex-traction module is introduced to correct the low-resolution sampling results, thus enhancing local detail expression while preserving the accuracy of multi-scale geometric structures. Then, a dynamic feature ex-traction module is introduced to capture long-range edge features, helping to alleviate jaggedness and de-tail loss caused by sparse information. Experimental results show that the proposed method achieves ex-cellent performance, with the average CD decreased by 3.5% compared with HRE-NDC and the F-Score in-creased by an average of 3.4% relative to HRE-NDC on the ABC and Thingi10k datasets. It outperforms existing methods in both visual and quantitative evaluations, effectively reducing edge jaggedness and surface irregularities, and significantly improving the overall quality of mesh reconstruction.