Abstarct: In recent years there has been a lot of attention in Feature selection. Therefore feature Selection and reduction of data dimension is very important in Pattern recognition. In real-time systems, feature reduction increases the speed of processing and needs less memory to store information. In the context of optimization algorithms, especially Particle swarm optimization (PSO) has been considerable. So that using PSO and choosing the appropriate method to calculate the PSO fitness can be reduced a significant amount of feature vectors to achieve acceptable results efficiently. In this thesis the features of Persian handwritten digits are extracted by combining the gradient histogram and characteristic loci methods. Also, we proposed the Binary particle swarm optimization (BPSO) algorithm by appropriate fitness function to select the important features. In order to digit recognition the features that selected in the proposed method classified by Support Vector Machine (SVM). The proposed method applied to HODA databaxse. The results are shown 99.40% accuracy without reducing the features (400 features) and 99.11% accuracy with reducing the features (193 features). Also, ORL face databaxse, FVC2004 Fingerprint databaxse and UCI databaxse are used to verify the performance of system. In comparison to other research results we show that the proposed method has a good performance in feature selection.