In this paper, we investigate how the positioning and size of input and reference patterns affect the use of multiple reference patterns in a classical Joint Transform Correlator (JTC), and we produce an optical system using a JTC with multiple reference patterns. This system is applied to feature extraction from handwritten alphabetical letters, using four linear elements aligned vertically, horizontally, and obliquely at angles of ±45 degrees as features. The extracted features were input into a three-layer neural network to decide which letter was input. This system achieved an average recognition rate of 88.8% for 30 sets of handwritten alphabetical letters.