An optical neural network model and a key device for its optical implementation are proposed. The model, named the vector-feature-extracting optical neural network, can correctly recognize hand-written letters and can easily be implemented in optics. The key device, named a feature-extracting optical neuron device, can selectively extract specific line segments included in an optical input pattern. In this paper the structure and recognition process of the vector-feature-extracting optical neural network are shown in detail. The function of the feature-extracting optical neuron device is experimentally shown. In addition, the optical system based on the vector-feature-extracting optical neural network is described.