For　technical　problems　of　low　accuracy　of　automatic　decoding　for　Morse　signal,　an　automatic　decoding　method　for　time-frequency　spectrum　of　manual　or　mechanical　Morse　signal　is　put　forward,　which　based　on　time　frequency　analysis　method　and　machine　learning　technology.　It　generates　time　frequency　image　based　on　STFT,　which　used　for　extraction　of　Morse　signal　based　on　adaptive　image　enhancement　later.　K-means　clustering　algorithm　have　been　introduced　to　identify　the　dots,　dashes　and　interval　between　them.　Error-correction　algorithm　put　forward　to　improve　the　accuracy　of　decoding.　Simulation　experiment　and　engineering　practice　on　Morse　signal　demonstrate　the　effectiveness　and　feasibility　of　this　algorithm.