The invention discloses a face portrait synthesis method based on deep coupling self-coding, which mainly solves the problem that in the prior art, when the training data is insufficient, a synthesized face portrait lacks part of specific identity information. The implementation scheme comprises respectively training a face photo coupling auto-encoder, a face portrait coupling auto-encoder and a single-layer neural network formed by the coupling auto-encoders, using the single-layer neural network to find a non-linear mapping relation between the face photo and the face portrait, putting the face photo into the trained coupling auto-encoder to generate a hidden training photo block with specific information, representing the test photo candidate block set by a training photo block and a weight matrix in a linear combination manner, obtaining a face portrait candidate block set through the similarity hypothesis, and fusing the human face components in the human face portrait candidate block set to form a human face portrait. According to the invention, the synthesized face portrait has more specific information, is cleaner and more realistic, and can be used for converting a figurephoto into a figure portrait in digital entertainment.