基于改进Deeplab v3+的服装图像分割网络

在服装图像分割领域，现有算法存在服装边缘分割粗糙、分割精度差和服装深层语义特征提取不够充分等问题。将Coordinate Attention机制和语义特征增强模块（SFEM）嵌入到语义分割性能较好的Deeplab v3+网络，设计一种用于服装图像分割领域的CA_SFEM_Deeplab v3+网络。为了加强服装图像有效特征的学习，在Deeplab v3+网络的主干网络resnet101中嵌入Coordinate Attention机制，并将经过带空洞卷积池化金字塔网络的特征图输入到语义特征增强模块中进行特征增强处理，从而提高分割的准确率。实验结果表明，CA_SFEM_Deeplab v3+网络在DeepFashion2数据集上的平均交并比与平均像素准确率分别为0.557、0.671，相较于Deeplab v3+网络分别提高2.1%、2.3%，其所得分割服装轮廓更为精细，具有较好的分割性能。

Clothing Image Segmentation Network Based on Improved Deeplab v3+

To solve the problems of rough clothing edge segmentation, unsatisfactory segmentation accuracy, and insufficient deep semantic feature extraction in clothing image segmentation, the Coordinate Attention(CA) mechanism and Semantic Feature Enhancement Module(SFEM) are embedded into the Deeplab v3+ network, whichfeatures good semantic segmentation performance, and a CA_SFEM_Deeplab v3+ network is proposed for clothing image segmentation in this study.To strengthen the learning of effective features in clothing images, the CA mechanism module is embedded into resnet101, which is the backbone network of the Deeplab v3+ network, and the feature map after convolution pooling is performed on a pyramid with holes is input into the SFEM for feature enhancement.Consequently, the segmentation accuracy improved.Experimental results show that the mean Intersection over Union(mIoU) and Mean Pixel Accuracy(MPA) of the CA_SFEM_Deeplabv3 + network are 0.557 and 0.671, respectively, in the DeepFashion2 dataset, which are 2.1% and 2.3% higher than those of the Deeplab v3 + network, respectively.Compared with the Deeplab v3+ network, the proposedCA_SFEM_Deeplab v3+offersa finer segmentation of the clothing contour and better segmentation performance.