| 引用本文: | 汤维杰,方 挺,韩家明,袁东祥.基于轻量级 MobileNetV2-DeeplabV3+的棒材分割方法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2024，41（3）：66-71 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 针对当前语义分割模型为提升像素分割精度,不断增加算法复杂度,导致模型出现参数量大,耗时长,难以
部署至工业现场等问题,提出一种基于轻量级 MobileNetV2-DeeplabV3+模型的棒材分割算法。 算法为平衡像素分
割精度、模型参数量和算法检测速度,在原网络基础上做出一系列改进:将原有的 Xception 主干网络替换为轻量级
MobileNetV2 网络以降低模型参数量与计算复杂度;在空洞空间金字塔池化(ASPP)模块基础上密集连接各空洞卷
积以获得更大的感受野,更加密集的像素采样,并扩大输出特征覆盖的语义信息;使用深度可分离卷积(DSConv)
替代 ASPP 模块中的标准卷积进一步降低模型的计算复杂度;此外,引入有效通道注意力(ECA)模块聚焦目标边
缘特征,增强特征图通道信息提取的效果。 实验表明:改进后的模型在棒材数据集下平均交并比( MIOU) 为
89. 37%,平均像素精度(MPA)为 94. 57%,帧率(FPS)为 33. 09 帧/ s,模型参数量为 33. 6 M。 与 U-net、M-PSPNet、
M-DeeplabV3+等模型相比,改进后算法的 MIOU 值与 MPA 值略低于最佳值,但仍处于较高水准,模型参数量小,
FPS 值得到较大提升。 实验表明:改进后的算法能较好地平衡分割精度和算法实时性,能满足部署至工业现场的
需求。 |
| 关键词: 语义分割 DeepLabv3+模型 轻量级 棒材 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Bar Segmentation Method Based on Lightweight MobileNetV2-DeeplabV3+ |
|
TANG Weijie FANG Ting HAN Jiaming YUAN Dongxiang
|
|
School of Electrical and Information Engineering Anhui University of Technology Anhui Maanshan 243000 China
|
| Abstract: |
| In order to improve the pixel segmentation accuracy of the current semantic segmentation model the algorithm
complexity continues to increase resulting in a large number of parameters time-consuming and difficulty in deploying
to industrial sites. A bar segmentation algorithm based on the lightweight MobileNetV2-DeeplabV3+ model was proposed.
The algorithm made a series of improvements based on the original network in order to balance the pixel segmentation
accuracy the number of model parameters and the detection speed of the algorithm. The original Xception backbone
network was replaced with a lightweight MobileNetV2 network to reduce the number of model parameters and
computational complexity. On the basis of the Atrous Spatial Pyramid Pooling ASPP module the atrous convolutions
were densely connected to obtain a larger receptive field and denser pixel sampling and to enlarge the semantic
information covered by the output features. The computational complexity of the model was further reduced by using deep
separable convolution DSConv instead of the standard convolution in the ASPP module. In addition an effective
channel attention ECA module was introduced to focus on the target edge features and enhance the effect of channel
information extraction in the feature maps. The experiment showed that the improved model achieved a mean intersection over Union MIOU of 89. 37% a mean pixel accuracy MPA of 94. 57% a frame rate of 33. 09 frames per second
FPS and a model parameter size of 33. 6 M on the bar dataset. Compared with the models of U-net MPSPNet and MDeeplabV3+ the MIOU and MPA values of the improved algorithm were slightly lower than the best values but still at a
high level with a small number of model parameters and a significant increase in FPS value. The example shows that the
improved algorithm can better balance the segmentation accuracy and the real-time performance of the algorithm and can
meet the needs of deployment to industrial sites. |
| Key words: semantic segmentation DeepLabv3+ model lightweight bar |
