Lots of researches have been done for face recognition by using visible (VIS) spectrum images and great success has been achieved in this field. Despite the achievements, face recognition by VIS spectrum images in uncontrolled optical conditions is already a challenging issue. In such conditions, using different types of imaging specially infrared imaging sensors can help to improve the efficiency of face recognition system in VIS spectrum. Recently using near infrared (NIR) spectrum images has been considered as an important information source in uncontrolled optical conditions. The purpose of this thesis is to improve the efficiency of face recognition system in VIS spectrum by use of NIR spectrum images. Following the purpose a method based on deep learning algorithms named coupled convolutional autoencoder (CPCAE) has been suggested. In this method two convolutional autoencoder neural networks (CAE) have been used for learning compact representations of NIR and VIS spectrum images. As the CAE neural network can only learn compact representations of face images related to one domain, it can not model the relationships between NIR and VIS images. Thus a vector weight has been used to learn a mapping of compact representations of NIR spectrum images to compact representations of VIS spectrum images. The efficiency of the suggested method on CASIA NIR -VIS 2.0 database images has been examined and the accuracy of heterogeneous face images recognition by combination of VIS and NIR images has been studied. Laboratory results show that the average accuracy of recognizing NIR spectrum image is 61/52 percent that shows the advantage of the suggested method to the same method introduced in this field. The average accuracy of recognizing VIS spectrum images has been calculated 62/23 percent. Keywords: 1. Deep learning, 2. Heterogeneous face recognition 3. VIS spectrum, 4. NIR spectrum, 5. Matching.