The　diagnosis　of　high-impedance　fault　(HIF)　is　a　critical　challenge　due　to　the　presence　of　faint　signals　that　exhibit　distortion　and　randomness.　In　this　study,　we　propose　a　novel　diagnostic　approach　for　HIF　based　on　semantic　segmentation　of　the　signal　envelope　(SE)　and　Hilbert　marginal　spectrum　(HMS).　The　proposed　approach　uses　1DUNet　to　identify　the　transient　process　of　potential　fault　events　in　zero-sequence　voltage　to　judge　fault　inception.　Longer　timescale　zero-sequence　voltage　is　then　used　to　extract　SE　and　HMS,　representing　HIF　distortion　and　randomness　characteristics.　These　features　are　transformed　into　images,　and　ResNet18　is　employed　to　detect　the　presence　of　HIF.　An　industrial　prototype　of　the　proposed　approach　has　been　implemented　and　validated　in　a　10　kV　test　system.　The　experimental　results　indicate　that　the　proposed　approach　outperforms　the　comparison　by　a　significant　margin　regarding　triggering　deviation　and　detection　accuracy,　particularly　in　resonant　distribution　networks.