基于编解码网络的化工驱油图像分割方法 |
点此下载全文 |
引用本文:翟钰杰 1,2 ,尚佳童 1,2 ,张 栋 1,2 ,赵伟强 3 ,雷 涛 1,2.基于编解码网络的化工驱油图像分割方法[J].计算技术与自动化,2022,(3):88-93 |
摘要点击次数: 310 |
全文下载次数: 0 |
|
|
中文摘要:基于深度学习的驱油图像分割是驱油率分析计算的关键步骤,相对于其他方法,基于U形结构的全卷积神经网络(Fully Convolutional Networks,FCN)在许多不同的图像任务中取得了显著的成效。但在这些结构中,标准卷积具有固定的感受野,不能提取到不同尺度的信息,丢失细节特征,导致分割精度较差,边缘模糊;其次主流模型对噪声的鲁棒性较差,并且信息之间往往会存在冗余,在融合阶段中未能对关键信息有效提取。为解决上述问题,在U-net的基础上引入多尺度信息提取和空间注意力融合并将其集成为多尺度信息提取融合(MIEF)模块,通过提取不同的尺度信息有效地保留了图像细节特征,之后通过空间注意力动态地融合提取到的多尺度信息,增强抗干扰性,实现网络对多尺度信息的有效利用。通过在驱油数据集上进行实验,准确率和MIoU相较于原始U-net网络分别提高了2.55%和1.24%,并与其他方法进行对比,验证了该方法的可行性和有效性。 |
中文关键词:图像语义分割 编解码网络 多尺度信息 空间注意力 |
|
Chemical Oil Displacement Image Segmentation Method Based on Codec Network |
|
|
Abstract:Oil displacement image segmentation based on deep learning is the key step of oil displacement rate analysis and calculation. Compared with other methods, the full convolution neural network (FCN) based on u-structure has achieved remarkable results in many different image tasks. However, in these structures, the standard convolution has a fixed receptive field, can not extract information of different scales, and loses detailed features, resulting in poor segmentation accuracy and blurred edges; Secondly, the mainstream model has poor robustness to noise, and there is often redundancy between information, so it can not effectively extract the key information in the fusion stage. In order to solve the above problems, based on U-net, this paper introduces multi-scale information extraction and spatial attention fusion, and integrates them into multi-scale information extraction and fusion (MIEF) module. By extracting different scale information, the image detail features are effectively retained, and then the extracted multi-scale information is dynamically fused through spatial attention to enhance anti-interference, Realize the effective utilization of multi-scale information in the network. Through experiments on oil displacement data sets, the accuracy and Miou are improved by 2.55% and 1.24% respectively compared with the original U-net network. Compared with other methods, the feasibility and effectiveness of this method are verified. |
keywords:image semantic segmentation codec network multi-scale information spatial attention |
查看全文 查看/发表评论 下载pdf阅读器 |
|
|
|