基于剪枝与知识蒸馏的轻量级人像抠图方法  

Lightweight Human Matting Method Based on Pruning and Knowledge Distillation

作　　者：程航[1] 徐树公 Cheng Hang

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《工业控制计算机》2024年第1期91-93,共3页Industrial Control Computer

摘　　要：近年来,随着人像抠图技术的广泛应用,对其实时性与精度的要求也逐渐提高。现有轻量级方法在精度上难以得到保证,而精度更高的方法往往使用较大的网络结构,无法满足实时性需求。为了解决这一问题,提出了一种基于剪枝与知识蒸馏的轻量级人像抠图方法。该方法首先通过网络剪枝来获得一个轻量级学生网络结构,随后使用该学生网络进行知识蒸馏。实验证明,该方法可以在保证模型精度的前提下有效减少参数量和推理耗时,且相较现有的轻量级人像抠图方法具有更少的参数量和更高的精度。In recent years,with the wide application of human matting technology,the requirements for its realtime performance and accuracy have also increased gradually.Existing lightweight methods are difficult to guarantee accuracy,while higher-precision methods often use larger network structures that cannot meet realtime requirements.In order to solve this problem,a lightweight human matting method based on pruning and knowledge distillation is proposed.This method first obtains a lightweight student network structure through network pruning,and then uses the student network for knowledge distillation.Experiments show that this method can effectively reduce parameter quantity and inference time while ensuring model accuracy,and has fewer parameters and higher accuracy than existing lightweight human matting methods.

关 键 词：神经网络 图像抠图 网络剪枝 知识蒸馏 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]