Binocular Vision Localization and Measurement Based on Lightweight HRNet

doi:10.3969/j.issn.1004-132X.2026.01.021

Abstract

Abstract:

Aiming at the problems of low efficiency in binocular vision measurements based on feature point detection and high computational complexity of neural networks， a binocular vision localization and measurement method was proposed based on a lightweight HRNet. The lightweight HRNet was built upon the original HRNet by replacing the convolutional modules to reduce the number of parameters， introducing Transformer to extract global image features， and employing a multi-level upsampling fusion strategy to capture the multi-scale feature information. Compared with the original HRNet model， the lightweight HRNet reduces model parameters by 95.40%， while computational loads and normalized mean errors are decreased by 94.27% and 6.25% respectively. In terms of 3D measurement， the relative errors of the method combining lightweight HRNet with binocular vision reache 0.256%， enabling high-precision detection on hardware with low computational power.

Key words: binocular vision, high resolution net （HRNet）, lightweight, landmark detection, measurement

摘要：

针对基于特征点检测的双目视觉测量效率低、神经网络计算复杂度高等问题，提出了基于轻量化高分辨率网络（HRNet）的双目视觉定位与测量方法。轻量化HRNet以HRNet为基准，先替换卷积模块、缩减参数量，再引入Transformer提取全局图像特征，最后使用多级上采样融合策略捕获多尺度特征信息。与原HRNet模型相比，轻量化HRNet模型参数减少95.40%，计算量、归一化平均误差分别减小94.27%和6.25%；三维测量上，轻量化HRNet与双目视觉结合方法的相对误差达到0.256%，能在低算力硬件上实现高精度检测。

关键词: 双目视觉, 高分辨率网络, 轻量化, 关键点检测, 尺寸测量

CLC Number:

TP391.41

REN Jiaqi, XU Sixiang, DONG Binhui, TANG Ao, SONG Yuchen. Binocular Vision Localization and Measurement Based on Lightweight HRNet[J]. China Mechanical Engineering, 2026, 37(1): 201-208.

任加琪, 许四祥, 董宾卉, 汤澳, 宋昱宸. 基于轻量化高分辨率网络的双目视觉定位与测量[J]. 中国机械工程, 2026, 37(1): 201-208.

Figures/Tables 15

Fig.1 General framework for positioning and measurement

Fig.2 Schematic diagram of continuous casting billet model

Fig.3 Partial dataset image

Fig.4 Lightweight HRNet continuous casting billet key point detection network architecture

Fig.5 Shuffle block structure

Fig.6 Depth separable convolution schematic

Fig.7 Transformer layer architecture

Fig.8 Fusion of feature maps at different resolutions

Fig.9 Structure diagram of the SCConv module

Tab.1 Performance evaluation and comparison of algorithms on datasets

算法	U-Net	Transunet	HRNet	本文算法
N_p/10⁶	5.70	93.20	22.65	1.04
N_F/（10⁹s $- 1$ ）	25.73	59.01	18.34	1.05
E_mse/mm	3.56	2.15	2.08	2.01
E_nme/%	0.76	0.50	0.48	0.45
T/ms	86.21	102.21	116.27	62.50

Tab.1 Performance evaluation and comparison of algorithms on datasets

算法	U-Net	Transunet	HRNet	本文算法
N_p/10⁶	5.70	93.20	22.65	1.04
N_F/（10⁹s $- 1$ ）	25.73	59.01	18.34	1.05
E_mse/mm	3.56	2.15	2.08	2.01
E_nme/%	0.76	0.50	0.48	0.45
T/ms	86.21	102.21	116.27	62.50

Fig.10 Experimental visual results

Tab.2 Measurement results of cuboid continuous casting billet sizes based on the proposed algorithm

编号	边	真实尺寸/mm	测量尺寸/mm		绝对误差/mm		相对误差/%
编号	边	真实尺寸/mm	均值	最优值	均值	最优值	均值	最优值
1	长	300.340	299.570	301.012	0.770	0.670	$-$ 0.256	0.223
1	宽	68.300	70.161	68.181	1.861	0.120	2.725	$-$ 0.175
2	长	301.900	304.095	298.969	2.195	1.929	0.727	$-$ 0.639
2	宽	31.200	33.091	31.146	1.891	0.054	6.062	$-$ 0.172
3	长	204.120	204.477	203.991	0.357	0.128	0.175	$-$ 0.063
3	宽	30.580	32.345	30.891	1.769	0.312	5.784	1.019
4	长	262.680	261.626	263.008	1.054	0.328	$-$ 0.401	0.125
4	宽	100.960	100.673	101.368	0.287	0.408	$-$ 0.285	0.404

Tab.2 Measurement results of cuboid continuous casting billet sizes based on the proposed algorithm

编号	边	真实尺寸/mm	测量尺寸/mm		绝对误差/mm		相对误差/%
编号	边	真实尺寸/mm	均值	最优值	均值	最优值	均值	最优值
1	长	300.340	299.570	301.012	0.770	0.670	$-$ 0.256	0.223
1	宽	68.300	70.161	68.181	1.861	0.120	2.725	$-$ 0.175
2	长	301.900	304.095	298.969	2.195	1.929	0.727	$-$ 0.639
2	宽	31.200	33.091	31.146	1.891	0.054	6.062	$-$ 0.172
3	长	204.120	204.477	203.991	0.357	0.128	0.175	$-$ 0.063
3	宽	30.580	32.345	30.891	1.769	0.312	5.784	1.019
4	长	262.680	261.626	263.008	1.054	0.328	$-$ 0.401	0.125
4	宽	100.960	100.673	101.368	0.287	0.408	$-$ 0.285	0.404

Tab 3 Measurement results of each algorithm

算法	1号连铸坯长边			3号连铸坯短边			运行时间/s
算法	测量尺寸/mm	真实尺寸/mm	相对误差/%	测量尺寸/mm	真实尺寸/mm	相对误差/%	运行时间/s
SIFT	305.680	300.340	1.778	27.091	30.580	$-$ 11.409	19.918
ORB	295.340		$-$ 1.663	25.836		$-$ 15.513	0.505
改进ORB^［3］	298.773		$-$ 0.520				0.640
改进KAZE^［4］	301.431		0.363	28.305		7.440	3.556
Transunet*深度学习^［19］	299.844		$-$ 0.165	32.807		7.284	0.105
本文算法	301.012		$-$ 0.256	32.345		5.784	0.063

Tab 3 Measurement results of each algorithm

算法	1号连铸坯长边			3号连铸坯短边			运行时间/s
算法	测量尺寸/mm	真实尺寸/mm	相对误差/%	测量尺寸/mm	真实尺寸/mm	相对误差/%	运行时间/s
SIFT	305.680	300.340	1.778	27.091	30.580	$-$ 11.409	19.918
ORB	295.340		$-$ 1.663	25.836		$-$ 15.513	0.505
改进ORB^［3］	298.773		$-$ 0.520				0.640
改进KAZE^［4］	301.431		0.363	28.305		7.440	3.556
Transunet*深度学习^［19］	299.844		$-$ 0.165	32.807		7.284	0.105
本文算法	301.012		$-$ 0.256	32.345		5.784	0.063

Tab.4 Ablation experiment results of simplified model structural

方法	N_p/10⁶	N_F/（10⁹s $- 1$ ）	$E n m e$ /%	T/ms
HRNet	22.64	18.34	0.48	116.27
HRNet-1	12.90	12.20	0.50	80.00
HRNet-2	10.08	9.93	0.52	60.61
HRNet-3	5.98	7.42	0.56	55.25

Tab.4 Ablation experiment results of simplified model structural

方法	N_p/10⁶	N_F/（10⁹s $- 1$ ）	$E n m e$ /%	T/ms
HRNet	22.64	18.34	0.48	116.27
HRNet-1	12.90	12.20	0.50	80.00
HRNet-2	10.08	9.93	0.52	60.61
HRNet-3	5.98	7.42	0.56	55.25

Tab.5 Ablation experiment results of each module

LT	TF	SC	N_p/10⁶	N_F/（10⁹s $- 1$ ）	$E n m e$ /%
			5.98	7.42	0.56
√			0.67	1.78	0.62
√	√		0.94	1.78	0.54
√		√	0.77	1.05	0.56
√	√	√	1.04	1.05	0.45

Tab.5 Ablation experiment results of each module

LT	TF	SC	N_p/10⁶	N_F/（10⁹s $- 1$ ）	$E n m e$ /%
			5.98	7.42	0.56
√			0.67	1.78	0.62
√	√		0.94	1.78	0.54
√		√	0.77	1.05	0.56
√	√	√	1.04	1.05	0.45

References 19

[1]	许四祥，陈富强，高培青，等. 一种去除板坯毛刺的系统： CN102935547B［P］. 2014-10-15.
	Xu Sixiang， Chen Fuqiang， Gao Peiqing， et al. System for Removing Slab Burrs： CN102935547B［P］. 2014-10-15.
[2]	XU Sixiang， DONG Chenchen， ZHOU Shuhua， et al. Binocular Measurement Method for the Continuous Casting Slab Model Based on the Improved BRISK Algorithm［J］. Applied Optics， 2022， 61（11）： 3019-3025.
[3]	周书华，许四祥，董晨晨，等. 基于局部信息熵和梯度漂移的双目视觉测量算法［J］. 激光与光电子学进展， 2023， 60（12）： 333-341.
	ZHOU Shuhua， XU Sixiang， DONG Chenchen， et al. Algorithm for Binocular Vision Measurements Based on Local Information Entropy and Gradient Drift［J］. Laser & Optoelectronics Progress， 2023， 60（12）： 333-341.
[4]	宋祥，许四祥，杨利法，等. 基于非线性扩散与高维M-SURF描述符的双目视觉测量方法［J］. 光电子·激光， 2024， 35（4）： 405-413.
	SONG Xiang， XU Sixiang， YANG Lifa， et al. Binocular Vision Measurement Method Based on Nonlinear Diffusion and High-dimensional M-SURF Descriptor［J］. Journal of Optoelectronics·Laser， 2024， 35（4）： 405-413.
[5]	谢阳，戴逸群，张超勇，等. 融合集成模型与深度学习的机床能耗识别与预测方法［J］. 中国机械工程， 2023， 34（24）： 2963-2974.
	XIE Yang， DAI Yiqun， ZHANG Chaoyong， et al. A Method for Identifying and Predicting Energy Consumption of Machine Tools by Combining Integrated Models and Deep Learning［J］. China Mechanical Engineering， 2023， 34（24）： 2963-2974.
[6]	BORRA S R， PREMALATHA B， DIVYA G， et al. Deep Hashing with Multilayer CNN-based Biometric Authentication for Identifying Individuals in Transportation Security［J］. Journal of Transportation Security， 2024， 17（1）： 4.
[7]	MASUDA Y， ISHIKAWA R， TANAKA T， et al. CNN-based Fully Automatic Mitral Valve Extraction Using CT Images and Existence Probability Maps［J］. Physics in Medicine & Biology， 2024， 69（3）： 035001.
[8]	SHELHAMER E， LONG J， DARRELL T. Fully Convolutional Networks for Semantic Segmentation［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2017， 39（4）： 640-651.
[9]	RONNEBERGER O， FISCHER P， BROX T. U-Net： Convolutional Networks for Biomedical Image Segmentation［M］∥Medical Image Computing and Computer-assisted Intervention—MICCAI 2015. Cham： Springer International Publishing， 2015： 234-241.
[10]	SUN Ke， XIAO Bin， LIU Dong， et al. Deep High-resolution Representation Learning for Human Pose Estimation［C］∥2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Long Beach， 2019： 5686-5696.
[11]	WANG Jian， LONG Xiang， CHEN Guowei， et al. U-HRnet： Delving into Improving Semantic Representation of High Resolution Network for Dense Prediction［J］．arXiv：2210.07140．
[12]	HOWARD A G， ZHU Menglong， CHEN Bo， et al. MobileNets： Efficient Convolutional Neural Networks for Mobile Vision Applications［J］． arXiv：1704.04861．
[13]	ZHANG Xiangyu， ZHOU Xinyu， LIN Mengxiao， et al. ShuffleNet： an Extremely Efficient Convolutional Neural Network for Mobile Devices［C］∥2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City， 2018： 6848-6856.
[14]	郑云飞，王晓兵，张雄伟，等. 基于金字塔知识的自蒸馏HRNet目标分割方法［J］. 电子学报， 2023， 51（3）： 746-756.
	ZHENG Yunfei， WANG Xiaobing， ZHANG Xiongwei， et al. The Self-distillation HRNet Object Segmentation Based on the Pyramid Knowledge［J］. Acta Electronica Sinica， 2023， 51（3）： 746-756.
[15]	VASWANI A， SHAZEER N， PARMAR N，et al. Attention is All You Need［C］∥Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach，2017：6000-6010．
[16]	CHEN Jieneng， LU Yongyi， YU Qihang，et al. TransUNet： Transformers Make Strong Encoders for Medical Image Segmentation［J］. arXiv：.
[17]	ZHANG Zhengyou. A Flexible New Technique for Camera Calibration［J］．IEEE Transactions on Pattern Analysis and Machine Intelligence，2000，22（11）：1330-1334．
[18]	LI Jiafeng， WEN Ying， HE Lianghua. 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.
[19]	李同谱，许四祥，施宇翔，等. 基于双目视觉与Transformer的连铸坯模型定位与测量［J］. 中南大学学报（自然科学版）， 2024， 55（4）： 1312-1322.
	LI Tongpu， XU Sixiang， SHI Yuxiang， et al. Continuous Casting Slab Model Positioning and Measurement Based on Binocular Vision and Transformer［J］. Journal of Central South University （Science and Technology）， 2024， 55（4）： 1312-1322.