人像自动提取与处理软件设计与实现  

Design and Implementation of an Automatic Portrait Extraction and Processing Software

作　　者：刘清雨 蒋翔 金鑫[1] LIU Qingyu;JIANG Xiang;JIN Xin(Beijing Electronic Science and Technology Institute,Beijing 100070,P.R.China)

机构地区：[1]北京电子科技学院,北京市100070

出　　处：《北京电子科技学院学报》2022年第2期51-60,共10页Journal of Beijing Electronic Science And Technology Institute

基　　金：北京高校“高精尖”学科建设项目(项目编号:20210070Z0401)。

摘　　要：人像提取是当前图像处理方向的前沿热点问题,具有广泛的应用前景,但面临着边缘噪声大、颜色相近的区域分割效果差等问题。本文针对该问题创造性地提出了BiSeGAN网络,研究并实现了一种人像自动提取与处理的方法。此方法以生成对抗网络(GAN)框架为基础,采用BiSeNet结构作为生成器网络进行训练,兼顾了空间分辨率和感受野,提高了模型的识别速度和精度。本文使用混合损失函数共同优化网络参数,帮助模型快速收敛,使生成器生成更真实的目标图像。本文实现了自动化提取并处理人像,提取的人像部分边缘分割清晰,信噪比高,有效地优化了人像提取的准确度。而后,基于百度飞浆在移动端部署实现。Currently,portrait extraction is a research focus of image processing and has a prospect of wide application.However,it faces several problems such as large edge noise and poor segmentation effect for the regions with similar colors.To address the problems,a BiSeGAN network is creatively proposed in this paper,and a method for automatic portrait extraction and processing is studied and implemented.The method is based on the Generative Adversarial Network(GAN)framework and the BiSeNet structure is utilized as the generator network for training.In this method,spatial resolution and receptive field are both considered to improve the recognition speed and the accuracy of model.In this paper,a hybrid loss function is used to jointly optimize the network parameters to converge the model quickly and to generate more realistic target images.The automatic portrait extraction and processing is practically realized.Extracted portraits have clear edge segmentation and high signal-to-noise ratio,which effectively optimizes the accuracy of portrait extraction.Then,based on Baidu PaddlePaddle,the portrait extraction is deployed on mobile terminals.

关 键 词：人像提取 BiSeGAN GAN BiSeNet 美颜滤镜 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]