Both　magnetic　Barkhausen　noise　(MBN)　and　tangential　magnetic　field　(TMF)　strength　can　be　applied　in　the　quantitative　prediction　of　surface　hardness　of　ferromagnetic　specimens.　The　prediction　accuracy　depends　on　the　selected　model　and　the　input　parameters　of　the　model.　In　this　study,　the　relationship　between　the　surface　hardness　of　12CrMoV　steel　plate　and　the　measured　MBN　and　TMF　signals　is　investigated　with　multivariable　linear　regression　(MLR)　model　and　BP　neural　network　technique.　A　comparative　study　between　the　MLR　and　BP　model　is　conducted.　The　external　validation　results　show　that　the　BP　model　utilizing　four　MBN　features　as　the　input　nodes　has　a　smaller　average　prediction　error　(3.7%)　than　that　of　the　MLR　model　(13.2%).　Features　extracted　from　the　MBN　and　TMF　signals　are　combined　together　as　the　input　parameters　of　the　BP　model　in　order　to　achieve　high　accuracy.　After　adding　two　more　TMF　features　into　the　input　nodes　of　the　BP　network,　the　external　validation　results　suggest　that　the　average　prediction　error　is　decreased　from　3.7　to　3.5%.