采用双注意力机制Deeplabv3+算法的服装图像分割  

Garment Image Segmentationn Using Dual Attention Mechanism Deeplabv3+ Algorithm

作　　者：赵乙 何嘉[1] ZHAO Yi;HE Jia(College of Computer science,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《成都信息工程大学学报》2022年第1期67-71,共5页Journal of Chengdu University of Information Technology

摘　　要：近年来服装时尚行业经济发展迅速,为了让用户选择服装和服装的设计更方便快捷,提高服装图像的分割效率尤为重要。目前的方法大多属于传统的分割方法,或者基于深度卷积神经网络(DCNN)。针对服装图像分割时易受背景、颜色、纹理等的影响,且服装的边缘分割不准确,基于Deeplabv3+算法提出了双注意力机制的方法识别分割服装图像,使用通道注意力机制和位置注意力机制构成名为CPAM的模块对Deeplabv3+网络进行改进。特征图经过多次下采样后再经过通道和位置注意力模块(CPAM)与ASPP模块并行,最后通过上采样得到预测图像。实验证明对不同场景的服装图像分割,加入CPAM模块的模型能更准确地将服装分割出来。In recent years,the garment fashion industry economy has developed rapidly.In order to make the user’s choice of clothing and clothing design more convenient and fast,it is particularly important to improve the efficiency of clothing segmentation.Most of the present methods are traditional segmentation methods or based on deep convolutional neural network(DCNN).For the clothing image segmentation task is easily affected by background,color,texture,etc.,and the clothing edge segmentation is not accurate,this paper proposes a method of dual-attention mechanism based Deeplabv3+algorithm to identify and segment clothing images.Channel attention mechanism and location attention mechanism are used to form a module named CPAM to improve Deeplabv3+network.After downsampling for several times,the feature image is parallel to the channel and position attention module(CPAM) and The ASPP module,and then the prediction image is obtained by upsampling.Finally,the experiment proves that the model with CPAM module can segment the clothing image more accurately in different scenes.

关 键 词：服装图像分割 DeepFashion2 Deeplabv3+ 语义分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]