中圖分類號:TP393 文獻標識碼:A 文章編號:1674-0033(2025)04-0039-08
Abstract:Inview of thedifferences in the recognitionabilitiesof BP neural networksand RBF networks in pixel-level data sample patern recognition,an improved BP neural network (adopting an adaptive learning rate and an additional momentum method) was constructed and compared with an RBF neural network. Three sets of experiments were designed: adding 5%~50% random noise,partially occluding letters,and introducing specific contour noise.The recognition performance was analyzed using confusion matrices.The results show that in the presence of random noise below 30% ,the recognition rates of both networksreach 100% .When the noise increases to 40% ,the recognition rate of the improved BP neural network drops to 84.61% ,while the RBF neural network still maintains 100% .Inthe face of partial occlusion,both networks perform well;however,under the interference of contour noise,therecognition rate of the RBF neural network (84.62%) issignificantly higher than that of the improved BP neural network (65.38% ).The results confirm that the RBF network has stronger generalization ability in complex noise scenarios and provides a better option for low- dimensional data recognition.
Key words:BP neural network; RBF neural network; pattern recognition
人工神經網絡于20世紀80年代蓬勃興起, 式的差異,神經網絡衍生出眾多類型。其中,BP迅速成為人工智能領域的研究焦點。根據連接方 (反向傳播)神經網絡(Backpropagation Neural
Network)作為核心分支,憑借強大的非線性函數映射能力,在諸多領域展現出卓越的應用價值。有關數字與字母識別的研究較多,如,CHEN等3將種群極值優化的競爭進化算法與分數階梯度下降機制相結合,構建出自適應分數階BP神經網絡PEO-FOBP,用于手寫體數字識別。LIU等4通過灰度化、二值化、平滑去噪等圖像預處理手段,實現了 85.88% 的識別準確率。張鳳南運用主成分分析(PCA)處理特征向量,并對比最小距離法、K近鄰法及BP神經網絡的數字識別效果。嚴雯怡借助PCA降維與特征加權融合策略,將MNIST手寫數字識別正確率提升至 96.8% 。RBF(徑向基函數)神經網絡(RadialBasis FunctionNeuralNetwork)是用于解決插值問題。后來,引入了自適應學習的算法,應用范圍擴展。RBF神經網絡具有快速的訓練速度和良好的泛化能力,在函數逼近、時間序列預測[等領域得到了廣泛應用。例如,齊晴設計了混合優化的RBF神經網絡車牌字符識別模型,并與CNN模型進行對比。魏巍等[13提出基于單形進化的RBF神經網絡訓練算法。盡管上述改進算法均致力于提升識別準確率,但不可避免地需要在模型復雜度、訓練效率、計算資源消耗及泛化性能之間進行權衡。本文通過三組對比試驗,系統探究改進BP神經網絡與RBF神經網絡在識別任務中的性能差異:第一組試驗通過添加不同強度的噪聲,利用混淆矩陣分析[4網絡的抗噪識別能力;……