摘要
针对槟榔去核工序中槟榔内核轮廓检测问题,提出一种基于语义分割的槟榔内核轮廓检测方法。分割模型以VGG16为基础网络,将全连接层替换为卷积层,增加了跳跃结构,将浅层特征经过采样后在同一尺度下与深层特征进行融合,并将常规卷积替换成扩张卷积,减少了学习参数,提升了分割模型的实时性,得到最终的FCN-Dilated-VGG-8s分割模型。该模型对槟榔图像分割的准确率达到98.79%,单张图像分割只需0.071 s,模型大小只有FCN-VGG-8s模型的37.5%。算法表现出良好的鲁棒性,实现了槟榔图像准确、快速分割。通过对分割完后的图像的边界提取,即可得到完整平滑的槟榔内核轮廓线。
A method for the detection of betel nut stones contour based on semantic segmentation is proposed. The segmentation model is based on VGG16,among it the fully connected layer is replaced with convolution layer,moreover a jump structure is introduced,and the lower level features are merged with higher level deep features under the same scale,in addition we replace the conventional convolution with the dilated convolution,it results in reduced learning parameters,the real-time performance of the segmentation model is improved. By above process,the final FCN-Dilated-VGG-8 s segmentation model is obtained. The accuracy of the model is 98.79% for betel nut image segmentation,and cost of single image segmentation is only 0.071 s,and the model complexity is only37.5% of the FCN-VGG-8 s model. The algorithm shows good robustness and achieves accurate and quick segmentation of betel nut images. By extracting the boundary of the segmented image,a complete and smooth betel nut outline can be obtained.
作者
朱泽敏
张东波
张莹
汪忠
ZHU Ze-min;ZHANG Dong-bo;ZHANG Ying;WANG Zhong(College of Information Engineering,Xiangtan University,Xiangtan,Hunan 411105,China;National Engineering Laboratory for Robot Visual Perception&Control Technology,Changsha,Hunan 410012,China)
出处
《计算技术与自动化》
2019年第4期105-112,共8页
Computing Technology and Automation
基金
国家自然科学基金资助项目(61602397)
湖南省自然科学基金资助项目(2017JJ2251,2017JJ3315)
湖南省重点学科建设项目
关键词
语义分割
边缘检测
深度学习
全卷积网络
扩张卷积
semantic segmentation
edge detection
deep learning
full convolution network
dilated convolution
