Journal of Applied Optics, Volume. 43, Issue 1, 67(2022)
Parts recognition method based on improved Faster RCNN
Figures & Tables(13)
Table 1. ResNet101 network parameters
View tableView in ArticleTable 1. ResNet101 network parameters
卷积类型 输出尺寸/像素 卷积尺寸与特征通道数 卷积层 122×122 7×7，64，步长=2 池化层 56×56 3×3，64，步长=2 卷积块 56×56 [1×1，64；3×3，64；1×1，256]×3 卷积块 28×28 [1×1，128；3×3，128；1×1，512]×4 卷积块 14×14 [1×1，256；3×3，256；1×1，1024]×23 卷积块 7×7 [1×1，512；3×3，512；1×1，2048]×3
Table 2. Experimental comparison of feature detection network
View tableView in ArticleTable 2. Experimental comparison of feature detection network
特征网络 召回率/% 准确率/% 单张检测时间/s VGG16 90.3 94.5 0.47 ResNet101 91.8 96.3 0.4
Table 3. Experimental comparison of networks with different characteristics
View tableView in ArticleTable 3. Experimental comparison of networks with different characteristics
特征网络 召回率/% 准确率/% VGG16 86.3 91.2 ZF-Net 85.6 88.8 ResNet50 89.2 92.3 ResNet101 91.7 94.2
Table 4. Model test results of different strategies
View tableView in ArticleTable 4. Model test results of different strategies
策略 特征网络 RPN改进 非极大 值抑制 多尺 度训练 召回率/% 准确率/% 1 VGG16 否 否 否 88.2 90.5 2 ResNet101 否 否 否 88.9 91.6 3 ResNet101 是 否 否 90.6 93.7 4 ResNet101 否 是 否 90.9 94.0 5 ResNet101 否 否 是 90.3 93.5 6 ResNet101 是 是 否 92.3 95.4 7 ResNet101 是 是 是 92.8 96.3
Table 5. Experimental results of different models
View tableView in ArticleTable 5. Experimental results of different models
网络模型 召回率/% 准确率/% 识别时间/s SSD 85.7 88.6 0.71 YOLOv3 89.5 91.3 0.45 Faster RCNN 89.3 91.5 0.61 本文方法 93.2 96.1 0.65
Tools
Get Citation
Copy Citation Text
Yi WANG, Zhengdong MA, Guanglin DONG. Parts recognition method based on improved Faster RCNN[J]. Journal of Applied Optics, 2022, 43(1): 67
Paper Information
Category: OE INFORMATION ACQUISITION AND PROCESSING
Received: Jul. 22, 2021
Accepted: --
Published Online: Mar. 7, 2022
The Author Email: MA Zhengdong (309113885@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20