Pupil Diopter Detection Approach Based on Improved YOLOv3
-
摘要: 针对瞳孔区域屈光度识别准确率低、检测效率低等问题,本文提出一种基于改进YOLOv3深度神经网络的瞳孔图像检测算法。首先构建用于提取瞳孔主特征的二分类检测网络YOLOv3-base,强化对瞳孔特征的学习能力。然后通过迁移学习,将训练模型参数迁移至YOLOv3-DPDC(Deep Pupil Diopter Classify),降低样本数据分布不均衡造成的模型训练困难以及检测性能差的难题,最后采用Fine-tuning调参快速训练YOLOv3多分类网络,实现了对瞳孔屈光度快速检测。通过采集的1200张红外瞳孔图像进行实验测试,结果表明本文算法屈光度检测准确率达91.6%,检测速度可达45 fps,优于使用Faster R-CNN进行屈光度检测的方法。Abstract: To address the problems of low diopter recognition accuracy and low detection efficiency in the pupil area, this paper proposes a pupil image detection algorithm based on an improved YOLOv3 deep neural network. First, a two-class detection network YOLOv3 base for extracting the main features of the pupil is constructed to strengthen the learning ability of the pupil characteristics. Subsequently, through migration learning, the training model parameters are migrated to YOLOv3-DPDC to reduce the difficulty of model training and poor detection performance caused by the uneven distribution of sample data. Finally, fine-tuning is used to quickly train the YOLOv3 multi-classification network to achieve accurate pupil diopter detection. An experimental test was performed using the 1200 collected infrared pupil images. The results show that the average accuracy of diopter detection using this algorithm is as high as 91.6%, and the detection speed can reach 45 fps; these values are significantly better than those obtained using Faster R-CNN for diopter detection.
-
Key words:
- pupil diopter detection /
- deep learning /
- YOLOv3 /
- multi-scale features /
- machine vision
-
表 1 瞳孔数据集C的划分
Table 1. The division of pupil data set C
Diopter grade Initial number of samples Number of
samples after
amplificationTraining samples Validation
samplesR1 120 2000 1600 400 R1 780 5000 4000 1000 R3 300 4000 3200 800 表 2 实验参数设置
Table 2. Experimental parameter settings
Parameter Parameter value Training image size/pixel 416×416 Number of iterations 9000 GPU quantity 1 IOU 0.6 The initial learning rate of
the previous stage0.001 The initial learning rate of
the latter stage0.0001 表 3 不同算法数据对比
Table 3. Data of different algorithms
Algorithm model Precision/% Recall/% Time/ms F1-score/% SSD 82.5 84.7 20.3 84.7 RetinaNet 84.7 82.4 32.6 82.4 Faster R-CNN 86.8 88.9 109.6 88.9 Our algorithm 91.6 90.5 22.3 90.5 表 4 传统算法数据统计
Table 4. Statistics of traditional algorithms
Algorithm model Accuracy/% Time/ms Traditional algorithm 71.5 130 表 5 本文算法得出的分类混淆矩阵
Table 5. The classification confusion matrix obtained by the algorithm in this paper
R1 R2 R3 R1 19 1 0 R2 1 17 2 R3 0 1 19 -
[1] 薛烽, 李湘宁. 一种基于图像处理的屈光度测量方法[J]. 光电工程, 2009, 36(8): 62-66, 74. https://www.cnki.com.cn/Article/CJFDTOTAL-GDGC200908017.htmXUE Feng, LI Xiangning. A method of diopter measurement based on image processing[J]. Photoelectric Engineering, 2009, 36(8): 62-66, 74. https://www.cnki.com.cn/Article/CJFDTOTAL-GDGC200908017.htm [2] 胡志轩. 基于红外图像的眼视力屈光度检测系统[D]. 武汉: 华中师范大学, 2018.HU Zhixuan. An Eye Vision Diopter Detection System Based On Infrared Image[D]. Wuhan: Center China Normal University, 2018. [3] 谢娟英, 侯琦, 史颖欢, 等. 蝴蝶种类自动识别研究[J]. 计算机研究与发展, 2018, 55(8): 1609-1618. https://www.cnki.com.cn/Article/CJFDTOTAL-JFYZ201808003.htmXIE Juanying, HOU Qi, SHI Yinghuan, et al. Research on automatic recognition of butterfly species[J]. Computer Research and Development, 2018, 55(8): 1609-1618. https://www.cnki.com.cn/Article/CJFDTOTAL-JFYZ201808003.htm [4] Redmon J, Divvala S, Girshick R, et al. You only look once: unified real-time object detection[C]//Proceedings of the Conference on Computer Vision and Pattern Recognition, 2016: 779-788. [5] Kaur P, Khehra B S, Pharwaha A. Deep transfer learning based multi way feature pyramid network for object detection in images[J]. Mathematical Problems in Engineering, 2021, 2021: 1-13. [6] Noh H, Hong S, Han B. Learning deconvolution network for semantic segmentation[C]//IEEE International Conference on Computer Vision, 2015: 5119-5127. [7] 刘智, 黄江涛, 冯欣. 构建多尺度深度卷积神经网络行为识别模型[J]. 光学精密工程, 2017, 25(3): 799-805. https://www.cnki.com.cn/Article/CJFDTOTAL-GXJM201703035.htmLIU Zhi, HUANG Jiangtao, FENG Xin. Building a multi-scale deep convolutional neural network behavior recognition model[J]. Optics and Precision Engineering, 2017, 25(3): 799-805. https://www.cnki.com.cn/Article/CJFDTOTAL-GXJM201703035.htm [8] 王娟, 刘嘉润, 李瑞瑞. 基于深度学习的红外相机视力检测算法[J]. 电子测量与仪器学报, 2019, 33(11): 36-43. https://www.cnki.com.cn/Article/CJFDTOTAL-DZIY201911005.htmWANG Juan, LIU Jiarun, LI Ruirui. Infrared camera vision detection algorithm based on deep learning[J]. Journal of Electronic Measurement and Instrument, 2019, 33(11): 36-43. https://www.cnki.com.cn/Article/CJFDTOTAL-DZIY201911005.htm [9] LIU Y P, JI X X, PEI S T, et al. Research on automatic location and recognition of insulators in substation based on YOLOv3[J]. High Voltage, 2020, 5(1): 62-68. doi: 10.1049/hve.2019.0091 [10] Murugan A, Nair S, Kumar K. Detection of skin cancer using SVM, random forest and KNN classifiers[J]. Journal of Medical Systems, 2019, 43(8): 683-686. [11] ZHU S G, DU J P, REN N, et al. Hierarchical-Based object detection with improved locality sparse coding[J]. Chinese Journal of Electronics, 2016, 25(2): 290-295. doi: 10.1049/cje.2016.03.015 [12] 温江涛, 王涛, 孙洁娣, 等. 基于深度迁移学习的复杂环境下油气管道周界入侵事件识别[J]. 仪器仪表学报, 2019, 40(8): 12-19. https://www.cnki.com.cn/Article/CJFDTOTAL-YQXB201908002.htmWEN Jiangtao, WANG Tao, SUN Jiedi, et al. Intrusion event identification of oil and gas pipeline perimeter in complex environment based on deep migration learning[J]. Chinese Journal of Scientific Instrument, 2019, 40(8): 12-19. https://www.cnki.com.cn/Article/CJFDTOTAL-YQXB201908002.htm [13] REN S, HE K, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149. [14] LIU W, Anguelov D, Erhan D, et al. SSD: single shot multibox detector[C]//Proceedings of European Conference on Computer Vision, 2016: 21-37.