摘要
针对边缘检测算法在衣物轮廓提取过程中易受噪声影响的问题,采用传统的图像处理方法和深度学习方法对衣物轮廓提取算法进行研究。传统的图像处理方法主要基于HSV模型和Canny算法,首先,通过HSV模型分割前景图像;然后,对二值化图像进行形态学处理;最后,利用Canny算法提取衣物轮廓,此方法能准确地提取大部分颜色的衣物轮廓。深度学习方法主要基于HED网络模型,针对HED网络模型输出边缘定位缺失和较为粗糙等问题,对HED网络模型进行改进,首先,去除第3、4阶段的池化层;然后,在特征融合阶段引入注意力机制;最后,融合Canny算法进行边缘细化。对比实验结果表明:HSV+Canny算法比Canny算法的ODS值和OIS值分别提升了13.16%和14.72%,检测速度有小幅提升;改进的HED网络模型比HED网络模型的ODS值和OIS值分别提升4.84%和3.97%的同时,检测速度持平。
In response to the problem of edge detection algorithms being easily affected by noise in the process of clothing contour extraction,traditional image processing methods and deep learning methods are used to study the clothing contour extraction algorithm.Traditional image processing methods are mainly based on the HSV model and Canny algorithm.Firstly,foreground images are segmented using the HSV model;Then,perform morphological processing on the binary image;Finally,the Canny algorithm is used to extract clothing contours,which can accurately extract clothing contours of most colors.The deep learning method is mainly based on the HED network model.To address the problems of missing and rough edge localization in the output of the HED network model,improvements are made to the HED network model.Firstly,the pooling layers in stages 3 and 4 are removed;Then,introduce attention mechanism in the feature fusion stage;Finally,the Canny algorithm is integrated for edge refinement.The comparative experimental results show that the HSV+Canny algorithm has improved the ODS and OIS values by 13.16%and 14.72%respectively compared to the Canny algorithm,with a slight increase in detection speed;The improved HED network model improves the ODS and OIS values by 4.84%and 3.97%respectively compared to the HED network model,while maintaining the same detection speed.
作者
黄启华
杜玉晓
HUANG Qihua;DU Yuxiao(School of Automation,Guangdong University of Technology,Guangzhou 510006,China)
出处
《自动化与信息工程》
2024年第4期1-9,17,共10页
Automation & Information Engineering
基金
国家自然科学基金(61976059,61640213)。
