改进DenseNet模型在工件表面粗糙度视觉检测中的应用
Application of Improved DenseNet Model to Visual Inspection of Workpiece Surface Roughness
周友行 1易倩 2杨文佳 2赵文杰2
作者信息
- 1. 湘潭大学机械工程与力学学院,湖南湘潭 411105;复杂轨迹加工工艺及装备教育部工程研究中心,湖南湘潭 411105
- 2. 湘潭大学机械工程与力学学院,湖南湘潭 411105
- 折叠
摘要
针对原始DenseNet模型检测工件表面粗糙度时间长、准确率较低的问题,结合卷积层滤波器注意力机制和批归一化层缩放系数提出一种工件表面粗糙度检测的深度学习模型.首先,利用注意力重要性值判定模块内的冗余通道.其次,在Dense Block模块内引入批归一化层缩放系数判别特征通道的重要程度.最后联合卷积层滤波器的注意力重要性值和批归一化层缩放系数裁剪冗余通道,实现模型剪枝.实验结果表明,原始 DenseNet模型检测工件表面粗糙度的准确率为91.875%,检测时间为 483 s.当剪枝率为 20%时,其检测效果最好,检测准确率为 96.875%,检测时间为 255 s.相比于原始DenseNet模型,改进后的DenseNet模型检测效果更好,在质量检测领域方面具有一定的应用前景.
Abstract
In order to solve the problem that the original DenseNet model with a long time and low accuracy to detect workpiece surface roughness,a deep learning model for workpiece surface roughness detection is proposed by combining the attention mechanism of convolutional layer filter and the scale coefficient of batch normalized layer.Firstly,the importance value of attention mechanism of convolution layer filter is used to determine the redundant channels in Dense Block module.Secondly,the scale coefficient of batch normalization layer was introduced into the Dense Block module of DenseNet model in order to distinguish the importance of feature channels.Finally,the attention importance value of convolution layer filter and scale coefficient of batch normalization layer are combined to crop redundant channels.Experimental results show that the accuracy of original DenseNet model is 91.875%,and the detection time is 483 s.When pruning rate is 20%,the detection accuracy was 96.875%,and detection time was 255 s.Comparing with the conventional model,the improved DenseNet model has better detection effect and larger application in the field of quality inspection.
关键词
粗糙度检测/深度学习/DenseNet/模型剪枝Key words
roughness measurement/deep learning/DenseNet/model pruning引用本文复制引用
基金项目
国家自然科学基金项目(52175254)
国家自然科学基金项目(51775468)
湖南省教育厅科学研究项目(20A505)
湖南省研究生科研创新项目(CX20210645)
湘潭大学研究生科研创新项目(XDCX2021B174)
出版年
2024
国家科技期刊平台
NSTL
万方数据