用于人像提取及半身像合成的生成对抗网络算法

Head Area Extraction and Portrait Synthesis Method Using GAN

利用生成对抗网络(generative adversarial network,GAN)进行标准上半身人像的合成,从普通人像照片中截取部分区域得到面部对齐后的标准化上半身合成图像,处理后的标准化人像实现了目标主体与背景的分离,可以有效地优化目标识别和分割算法的结果.图像的合成过程分为2个主要步骤,首先利用图像特征识别人脸并截取头部区域,然后以裁切后的头部区域为中心进行上半身人像的合成,得到人脸特征点及头部区域对齐后的上半身合成图像.该算法可以有效地从背景中分离人像区域,利用合成后的图像进行图像分割和评价,可以避免图像背景对于图像识别主体的干扰.通过自有数据集验证了该算法可以改善分割算法的精确度、召回率和F值,最终合成人脸图像的Facenet平均距离及标准差相比现有的人脸图像正则化算法均有减小,通过在CelebA及LFW等通用数据集上的验证测试,显示出算法具有良好的通用性和适应性,该算法可以广泛适用于人像照片的主体提取和人像合成,作为分割和识别等应用的前置步骤.

The paper described a general method for portrait synthesis using the generative adversarial network(GAN),which can generate a standard feature point aligned portrait image by cropping facial area from an in-the-wild photo.The main target of processed image is separated from background and the object detection and segmentation algorithm results are optimized.The processing pipeline includes two main parts:firstly,recognize head area using low-level hand-craft features;secondly,use the cropped area as the input of GAN to synthesis portrait image with facial feature aligned.This method can effectively extract facial parts of the image and avoid affection from the background pattern and objects,as well as enhance the facial segmentation of existing algorithms.The experimental results optimized the precision,recall and F-measure values of the existing segmentation algorithm,demonstrated in CelebA and LFW datasets,which are different from the self-made training dataset,and decreased the Facenet distance and standard deviation compared with the state-of-art face frontalization algorithms,showed well generalization ability and proved that this method can be widely used as preprocessing of image segmentation and portrait synthesis methods.