素描是艺术创作中一种快速的自由创作方式。它拥有着独特的艺术价值，是最早出现 的艺术形式之一，也是视觉创作的基本形式之一。在公安刑侦领域，素描有着独特的应用 价值。在公安机关在缉拿犯人时往往无法得到准确的犯人照片，需要通过对目击者的口述 进行素描画像，再与数据库中照片进行匹配。然而，素描和照片属于不同的图像域，基于 不同域间的图像搜索技术仍未成熟。一个可行的方法是将照片转换为素描进而构建人脸素 描数据库并基于此进行检索。目前，人脸素描合成已成为计算机视觉和机器学习领域的具 有挑战性的问题。 本文首先对已有的人脸照片到素描合成的算法进行了介绍。传统的素描合成算法主要 分为基于子空间学习的人脸素描合成方法，基于稀疏表示的人脸素描合成方法和基于贝叶 斯推理的人脸素描合成方法。然后，本文对深度学习模型进行了介绍，介绍了卷积神经网 络和生成对抗网络模型。通过生成对抗网络的生成器和鉴别器合成素描图像，并且对合成 结果进行优化。进而构建人脸素描合成的基本网络构架。 我们注意到素描的绘画过程是有时序性特征的，即画家们在进行素描绘制的过程中通 常以类似的顺序进行绘制。绘制一幅素描的过程可划分为固定的阶段。利用这种特性，本 文建立了一个新的数据库，请画家对人脸素描进行分步骤绘制，即每张素描的绘制分为 25 步。基于所构建的分阶段数据库，提出了一种多阶段的生成对抗网络的素描合成方法，并 使用其中关键的五步进行人脸素描合成。 本文对上述的算法进行了大量实验评估，分为主观视觉评估、客观质量评估、进而进 行分析和总结，最后我们归纳总结了各个方法的优劣。 

A sketch is a rapidly executed freehand drawing. It is also one of the earliest forms of arts. In literature, a sketch has its unique artistic value. In public security and law enforcement, sketchbased face recognition is utilized for. In case thesuspect’s photo is not available, the police have to ask an artist to draw a sketch according to the description from the witnesses. Then, sketchbased face recognition is utilized. However, since sketch and photo belong to different image domain, cross domain image retrieval technique is not mature up to now. An alternative way is to synthesis face sketch from a face photo, then perform sketch-based face recognition. Therefore, face sketch synthesis becomes vital and it is also challenging. In this paper, firstly, we introduce existing face sketch synthesis algorithm. Traditional sketch synthesis algorithms can be categorized into face methods based on subspace learning, sparse representation as well as Bayesian reasoning. Then, we introduce to synthesize face sketch via deep learning model, especially convolutional neural networks (CNN) and Generative Adversarial Networks (GAN).  Noticing that in the process of sketch drawing, artists usually draw in a similar order. The whole drawing procedure can be finished in several steps. In each step, artists draw the sketch with different emphasis. Therefore, we build a new face sketch database which takes advantages of the sketch drawing order. In the new database, a sketch is drawn with twenty-five steps represented by intermediate sketch drawing. Based on the database, we utilized a multi-stage GAN for sketch synthesis in which five steps are adopted. Finally, experiments are implemented with all the considered methods. Subjective visual analysis, objective image quality evaluation and subjective quality evaluation are performed in different ways. Sketch based face recognition experiments prove the effectiveness of the proposed method.