A Method of Multi-type Job Identification Based on Improved Faster-RCNN

doi:10.3969/j.issn.1004-132X.2026.05.023

Abstract

Abstract:

To solve the problems of missing workpiece detection and low accuracy of workpiece detection in traditional machine vision algorithm， an improved Faster-RCNN multi-type job identification method was proposed. Firstly， at the backbone network level， the original VGG16 was discarded， the Res2Net101 was adopted to strengthen the efficiency of network feature extraction. Secondly， in the feature fusion module， the SCConv was deeply embedded into the FPN， and then a new SCFPN architecture was constructed， and the architecture was organically integrated with Res2Net101 to realize the full extraction and refinement of multi-scale feature information. Finally， in the post-processing link， Soft-NMS algorithm was used to replace the traditional non maximum suppression（NMS） algorithm， effectively avoiding the false deletion of high coincidence candidate box， significantly improving the adaptability of the model in the workpiece occlusion scene， and greatly reducing the probability of missing detection. The results show that the proposed improved Faster-RCNN model significantly improves the performances compared with that of the original algorithm， the average precision reaches 93.2% and recall rate is as 91.8%. It may detect and identify all kinds of workpieces in complex environment.

Key words: workpiece detection, improved Faster-RCNN, feature pyramid network（FPN）, spatial and channel reconstruction convolution（SCConv）

摘要：

针对传统机器视觉算法在工件检测任务中出现工件漏检、工件检测准确率低的问题，提出了一种改进的Faster-RCNN多种类工件识别方法。首先在骨干网络层面摒弃原有的VGG16，选用Res2Net101强化网络特征提取效能；其次在特征融合模块中将空间和通道重建卷积（SCConv）深度嵌入特征金字塔网络（FPN），进而构建全新的SCFPN 架构，并将其与Res2Net101有机整合，以实现对多尺度特征信息的充分提取与精炼；最后在后处理环节，以Soft-NMS 算法取代传统非极大值抑制（NMS）算法，有效规避高重合度候选框的误删除现象，显著提高模型在工件遮挡场景下的适配能力，大幅降低漏检概率。研究结果表明：所提改进Faster-RCNN模型相比原算法在性能上有显著提高，平均精度达到93.2%，召回率达到91.8%，可在复杂环境下完成对各类工件的检测识别。

关键词: 工件检测, 改进Faster-RCNN, 特征金字塔网络, 空间和通道重建卷积

CLC Number:

TP391.4

MEI Qizhen, SUN Xuan. A Method of Multi-type Job Identification Based on Improved Faster-RCNN[J]. China Mechanical Engineering, 2026, 37(5): 1226-1235.

梅旗振, 孙旋. 基于改进Faster-RCNN的多种类工件识别方法[J]. 中国机械工程, 2026, 37(5): 1226-1235.

Figures/Tables 17

Fig.1 Overall framework of multiple types of workpiece identification and detection

Fig.2 Faster-RCNN structure diagram

Fig.3 Improved Faster-RCNN structure diagram

Fig.4 Structure diagram of Res2Net module and ResNet module

Fig.5 FPN structure diagram

Fig.6 SCFPN structure diagram

Fig.7 SCConv module structure diagram

Fig.8 SRU unit structure diagram

Fig.9 CRU unit structure diagram

Fig.10 Different types of job images

Tab.1 Number of marking boxes for various types of workpieces

类别	框数
轴承	4041
螺栓	5223
凸盘	1241
齿轮	4574
螺母	6266
弹簧	2704
总计	24 049

Fig.11 Broken line diagram of model loss value

Tab.2 Performance comparison of different feature extraction modules in detection of various types of workpieces

算法模型	平均精度值 mAP@50/%	召回率 $A R$ /%
Faster-RCNN（VGG16Net）	83.1	78.6
Faster-RCNN （MobileV2）	82.3	76.8
Faster-RCNN （ResNet50）	83.3	79.8
Faster-RCNN （ResNet101）	86.2	84.4
Faster-RCNN（Res2Net101）	86.9	85.5

Tab.2 Performance comparison of different feature extraction modules in detection of various types of workpieces

算法模型	平均精度值 mAP@50/%	召回率 $A R$ /%
Faster-RCNN（VGG16Net）	83.1	78.6
Faster-RCNN （MobileV2）	82.3	76.8
Faster-RCNN （ResNet50）	83.3	79.8
Faster-RCNN （ResNet101）	86.2	84.4
Faster-RCNN（Res2Net101）	86.9	85.5

Tab.3 Performance comparison of different improvement strategies in multi-type workpiece detection

模型	平均精度值 mAP@50/%	召回率 $A R$ /%	帧率/ （帧·s $- 1$ ）
Faster-RCNN+1+2+3+4	93.2	91.8	12
Faster-RCNN+1+3+4	92.0	89.9	12
Faster-RCNN+1+2+3	91.4	90.2	12
Faster-RCNN+1+3	90.6	88.3	12
Faster-RCNN+1+2	89.1	87.5	16
Faster-RCNN+3+4	84.7	81.0	9
Faster-RCNN+3	84.5	80.2	9
Faster-RCNN+2	83.9	79.8	10
Faster-RCNN+1	86.9	85.5	16
Faster-RCNN	83.1	78.6	10

Tab.3 Performance comparison of different improvement strategies in multi-type workpiece detection

模型	平均精度值 mAP@50/%	召回率 $A R$ /%	帧率/ （帧·s $- 1$ ）
Faster-RCNN+1+2+3+4	93.2	91.8	12
Faster-RCNN+1+3+4	92.0	89.9	12
Faster-RCNN+1+2+3	91.4	90.2	12
Faster-RCNN+1+3	90.6	88.3	12
Faster-RCNN+1+2	89.1	87.5	16
Faster-RCNN+3+4	84.7	81.0	9
Faster-RCNN+3	84.5	80.2	9
Faster-RCNN+2	83.9	79.8	10
Faster-RCNN+1	86.9	85.5	16
Faster-RCNN	83.1	78.6	10

Tab.4 Performance comparison of different model algorithms in multi-type workpiece detection

模型	平均精度值 mAP@50/%	召回率 $A R$ /%	帧率/ （帧·s $- 1$ ）
SSD	82.3	76.8	32
YOLOV5	89.1	82.4	56
YOLOV8	88.9	83.0	53
原始Faster-RCNN	83.1	78.6	10
改进Faster-RCNN	93.2	91.8	12

Tab.4 Performance comparison of different model algorithms in multi-type workpiece detection

模型	平均精度值 mAP@50/%	召回率 $A R$ /%	帧率/ （帧·s $- 1$ ）
SSD	82.3	76.8	32
YOLOV5	89.1	82.4	56
YOLOV8	88.9	83.0	53
原始Faster-RCNN	83.1	78.6	10
改进Faster-RCNN	93.2	91.8	12

Tab.5 Mean average precision value mAP@50 when different model algorithms are used to detect various types of workpieces

模型	轴承	螺栓	凸盘	齿轮	螺母	弹簧
SSD	84.9	81.2	94.8	75.8	84.3	72.3
YOLOV5	88.9	89.5	94.3	83.3	90.7	87.9
YOLOV8	88.4	89.1	94.2	83.4	90.9	87.5
原始Faster-RCNN	84.8	83.6	94.7	76.7	86.3	72.4
改进Faster-RCNN	92.5	90.6	98.6	92.1	92.3	92.6

Fig.12 Effect picture of workpiece detection

References 21

[1]	闫纪红，李柏林.智能制造研究热点及趋势分析［J］.科学通报，2020，65（8）：684-694.
	YAN Jihong， LI Bolin. Research Hotspots and Tendency of Intelligent Manufacturing［J］. Chinese Science Bulletin， 2020， 65（8）： 684-694.
[2]	齐二石，霍艳芳，刘洪伟. 面向智能制造的工业工程和精益管理［J］. 中国机械工程， 2022， 33（21）： 2521-2530.
	QI Ershi， HUO Yanfang， LIU Hongwei. Industrial Engineering and Lean Management for Smart Manufacturing［J］. China Mechanical Engineering， 2022， 33（21）： 2521-2530.
[3]	王耀南，陈铁健，贺振东，等. 智能制造装备视觉检测控制方法综述［J］. 控制理论与应用， 2015， 32（3）： 273-286.
	WANG Yaonan， CHEN Tiejian， HE Zhendong， et al. Review on the Machine Vision Measurement and Control Technology for Intelligent Manufacturing Equipment［J］. Control Theory & Applications， 2015， 32（3）： 273-286.
[4]	车录锋，周晓军，徐志农，等. 多传感器信息融合机器人工件识别［J］. 中国机械工程， 2000， 11（11）： 1266-1268.
	CHE Lufeng， ZHOU Xiaojun， XU Zhinong， et al. Multisensor Data Fusion for Robot Parts Recognition［J］. China Mechanical Engineering， 2000， 11（11）： 1266-1268.
[5]	韩博. 基于双目立体视觉的工件识别与定位关键技术研究［D］. 哈尔滨：哈尔滨工业大学， 2018.
	HAN Bo. Research on Key Technologies of Workpiece Recognition and Location Based on Binocular Stereo Vision［D］. Harbin： Harbin Institute of Technology， 2018.
[6]	OTSU N. A Threshold Selection Method from Gray-level Histograms［J］. IEEE Transactions on Systems， Man， and Cybernetics， 1979， 9（1）： 62-66.
[7]	RUTA A， LI Y， LIU X. Real-time Traffic Sign Recognition from Video by Class-specific Discriminative Features［J］. Pattern Recognition， 2010，43（1）：416-430.
[8]	KHAN J F， BHUIYAN S M A， ADHAMI R R. Image Segmentation and Shape Analysis for Road-sign Detection［J］. IEEE Transactions on Intelligent Transportation Systems， 2011， 12（1）： 83-96.
[9]	黄山. 基于机器视觉的机械臂分拣系统的研究［D］. 镇江：江苏科技大学， 2016.
	HUANG Shan. Research on the Robot Arm Sorting System Based on Machine Vision［D］. Zhenjiang：Jiangsu University of Science and Technology，2016.
[10]	郭庆梅，刘宁波，王中训，等. 基于深度学习的目标检测算法综述［J］. 探测与控制学报，2023，45（6）：10-20.
	GUO Qingmei， LIU Ningbo， WANG Zhongxun， et al. Review of Deep Learning Based Object Detection Algorithms［J］. Journal of Detection & Control， 2023， 45（6）： 10-20.
[11]	CUI G， ZHANG L. Improved Faster Region Convolutional Neural Network Algorithm for UAV Target Detection in Complex Environment［J］. Results in Engineering， 2024， 23： 102487.
[12]	KONG X， LI X， ZHU X， et al. Detection Model Based on Improved Faster-RCNN in Apple Orchard Environment［J］. Intelligent Systems with Applications， 2024， 21： 200325.
[13]	LIU B， ZHAO X， HU H， et al. Detection of Esophageal Cancer Lesions Based on CBAM Faster R-CNN［J］. Journal of Theory and Practice of Engineering Science， 2023， 3（12）： 36-42.
[14]	ZHANG W， ZHU Q， LI Y， et al. MAM Faster R-CNN： Improved Faster R-CNN Based on Malformed Attention Module for Object Detection on X-ray Security Inspection［J］. Digital Signal Processing， 2023， 139： 104072.
[15]	THOKE A， RAI S. Exploring Faster R-CNN Algorithms for Object Detection［C］∥2024 First International Conference on Software， Systems and Information Technology（SSITCON）. Tumkur， 2024： 1-5.
[16]	MAO W L， CHEN S H， HUANG Y T， et al. Indoor Scene Recognition Using ARM-based MobileNets Architectures［C］∥2023 Fourteenth International Conference on Ubiquitous and Future Networks（ICUFN）. Paris， 2023： 225-230.
[17]	YAN X， ZHANG Y， JIN Q. Chemical Process Fault Diagnosis Based on Improved ResNet Fusing CBAM and SPP［J］. IEEE Access， 2023， 11： 46678-46690.
[18]	GAO S， CHENG M， ZHAO K， et al. Res2Net： a New Multi-scale Backbone Architecture［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2021， 43（2）： 652-662.
[19]	LIN T Y， DOLLÁR P， GIRSHICK R， et al. Feature Pyramid Networks for Object Detection［C］∥2017 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Honolulu， 2017： 936-944.
[20]	LI J， WEN Y， HE L. SCConv： Spatial and Channel Reconstruction Convolution for Feature Redundancy［C］∥2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Vancouver， 2023： 6153-6162.
[21]	刘罡，侯恩翔，黄孙港，等. 多尺度特征融合增强的行人翻越护栏检测［J］. 电子测量技术， 2024， 47（14）： 127-138.
	LIU Gang， HOU Enxiang， HUANG Sungang， et al. Multi Scale Feature Fusion Enhanced Pedestrian Crossing Guardrail Detection［J］. Electronic Measurement Technology， 2024， 47（14）： 127-138.