基于CNN的工业钉类识别研究及优化策略
|
摘要
为了实现交换机外壳上钉类安装正确与否的精确检测,文章提出一种基于改进卷积神经网络的图像处理识别算法。首先对交换机外壳图像进行Canny算子提取边缘,用透视变换算法与模板形成统一尺寸,定位截取出样本,然后把不同环境下的样本输入Caffe框架下构建的深度卷积神经网络,CPU模式下训练识别模型,最后运用于实际工程中。测试分析结果表明,随着不同钉类样本数据量增加,网络结构不能完全满足检测需求,进而采用一种基于产品类型的不同模型训练优化策略。实验结果表明,改进的算法可以快速训练模型并且结合策略可使不同产品检测准确度达99%以上,有效提高了交换机外壳上钉类安装的检测精度。
To realize the accurate detection of the nails of the switch housing in the industry, this paper proposes an image processing recognition algorithm based on improved convolutional neural network. Firstly, the Canny operator extracts the edge of the switch shell image and the perspective transformation algorithm is used to form a uniform size with the template, and the samples are located and intercepted. Then, the samples in different environments are input into the depth convolution neural network constructed under the Caffe framework, and the recognition model is trained under CPU mode. Finally, the recognition model is applied to practical engineering. The test analysis results show that with the increase of the data volume of different nail samples, the network structure can not fully meet the detection requirements, and then a different model training optimization strategy based on product types is adopted. The experimental results show that the improved algorithm can quickly train the model and combine the strategy to make the detection accuracy of different products reach more than 99%, which effectively improves the detection accuracy of the nail installation on the switch housing.
To realize the accurate detection of the nails of the switch housing in the industry, this paper proposes an image processing recognition algorithm based on improved convolutional neural network. Firstly, the Canny operator extracts the edge of the switch shell image and the perspective transformation algorithm is used to form a uniform size with the template, and the samples are located and intercepted. Then, the samples in different environments are input into the depth convolution neural network constructed under the Caffe framework, and the recognition model is trained under CPU mode. Finally, the recognition model is applied to practical engineering. The test analysis results show that with the increase of the data volume of different nail samples, the network structure can not fully meet the detection requirements, and then a different model training optimization strategy based on product types is adopted. The experimental results show that the improved algorithm can quickly train the model and combine the strategy to make the detection accuracy of different products reach more than 99%, which effectively improves the detection accuracy of the nail installation on the switch housing.
引文
[1] 冯娟,万星.基于改进阈值分割与形态学的螺钉目标检测算法[J].组合机床与自动化加工技术,2016(10):120-123.
[2] 唐聪,凌永顺,杨华,等.基于深度学习物体检测的视觉跟踪方法[J].红外与激光工程,2018,47(5):0526001.
[3] 段振云,单紫薇,赵文辉,等.数字图像边缘的快速提取算法研究[J].组合机床与自动化加工技术,2017(12):12-14.
[4] 何援军.透视和透视投影变换——论图形变换和投影的若干问题之三[J].计算机辅助设计与图形学学报,2005,17(4):734-739.
[5] 徐一丁,杜慧敏,毛智礼,等.基于卷积神经网络的工件识别算法[J].组合机床与自动化加工技术,2018(4):42-45,50.
[6] 樊恒,徐俊,邓勇,等.基于深度学习的人体行为识别[J].武汉大学学报(信息科学版),2016,41(4):492-497.
[2] 唐聪,凌永顺,杨华,等.基于深度学习物体检测的视觉跟踪方法[J].红外与激光工程,2018,47(5):0526001.
[3] 段振云,单紫薇,赵文辉,等.数字图像边缘的快速提取算法研究[J].组合机床与自动化加工技术,2017(12):12-14.
[4] 何援军.透视和透视投影变换——论图形变换和投影的若干问题之三[J].计算机辅助设计与图形学学报,2005,17(4):734-739.
[5] 徐一丁,杜慧敏,毛智礼,等.基于卷积神经网络的工件识别算法[J].组合机床与自动化加工技术,2018(4):42-45,50.
[6] 樊恒,徐俊,邓勇,等.基于深度学习的人体行为识别[J].武汉大学学报(信息科学版),2016,41(4):492-497.