Principal　component　analysis　(PCA)　has　been　applied　widely　in　pattern　recognition.　Based　on　the　nonlinear　PCA　algorithm　and　subspace　pattern　recognition　method,　a　nonlinear　PCA　neural　network　model　of　signal　reconstruction　has　been　proposed　in　this　paper.　The　method　has　been　used　in　handwritten　digits　and　characters　recognition,　and　a　comparison　with　BP　neural　network　based　classifiers　has　been　made.　Some　satisfactory　results　have　been　obtained.　The　experiment　results　show　that　the　average　correct　identification　rate　of　our　method　is　up　to　94.74%　for　the　handwritten　digits,　and　91.03%　for　the　handwritten　characters.