Visual　classification　has　attracted　considerable　research　attention　in　the　last　few　decades.　This　paper　presents　an　approach　to　improve　the　hypergraph　structure　using　a　diffusion　ranking　method　and　employs　hypergraph　neural　networks　for　such　tasks.　Hyperedges　are　formed　by　connecting　a　set　of　vertices,　and　the　generation　of　the　hypergraph　structure　is　typically　task-dependent,　making　hypergraph　generation　methods　less　generalizable.　Previous　approaches　have　employed　implicit　hypergraphs　for　visual　tasks,　especially　in　visual　classification,　using　attribute-based　or　neighborhood-based　methods.　However,　these　approaches　cannot　fully　explore　the　global　manifold　structure　of　the　feature　space.　To　address　this　problem,　this　paper　proposes　a　novel　two-stage　hypergraph　construction　method,　which　fully　utilizes　the　ability　of　diffusion　process　to　capture　the　manifold　structure　of　feature　data.　Initially,　the　hypergraph　structure　is　constructed　using　a　neighbor-based　approach,　and　subsequently,　the　final　hypergraph　structure　is　obtained　by　ranking　scores　of　diffusion　processes　that　reflect　the　degree　of　vertex　similarity.　Comparing　with　state-of-the-art　hypergraph　neural　network　algorithms　on　classical　3D　visual　image　datasets,　experimental　results　demonstrate　the　effectiveness　of　the　proposed　algorithm.　©　2024　IEEE.