基于全尺度融合侧输出残差网络的骨架提取算法  

作　　者：莫莉莎 王斌[1] 

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134-135,169,共3页Industrial Control Computer

摘　　要：图像骨架是一种紧凑、直观的图像表示方法。目前,现有基于深度学习的二值像素图像骨架提取算法存在骨架断裂问题。提出了一种基于全尺度融合侧输出残差Unet网络(FFSR-Unet)的骨架提取算法,该算法通过融合编码器和解码器不同层级间的特征,实现了前景对象不同形状尺度的特征交互,同时采用阶梯式残差块以增强网络对深层与浅层语义的提取能力。该算法在Pixel SkelNetOn Challenge数据集上的F1-score可以达到0.8548,能够超越现有算法的提取结果。Image skeleton is a compact and intuitive image representation method.Currently,the existing skeleton extraction of binary pixel images based on depth learning faces the problem of skeleton breakage.This paper proposes a fusing full-scale side outputs residual unet network(FFSR-Unet)of skeleton extraction algorithm.This network achieves feature interaction of foreground objects with different shape scales by fusing features between different levels of encoders and decoders,and uses Stepwise-Resblock to enhance the network’s ability to extract deep and shallow semantics.On the Pixel SkelNetOn Challenge dataset,the F1-score obtained according to this network model can reach 0.8548,which is able to surpass the extraction results of existing algorithms.

关 键 词：骨架提取 深度学习 二值像素图像 全尺度融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]