基于改进YOLOv5的皮革瑕疵检测算法

doi:10.19678/j.issn.1000-3428.0064587

摘要/Abstract

摘要：

皮革瑕疵检测是工业皮革生产行业中质量控制的重要环节，针对工业皮革瑕疵在线检测中存在计算复杂度高、对小目标检测效果差、数据样本少等问题，提出一种基于注意力机制的轻量化皮革瑕疵检测算法GPC-YOLOv5。使用工业相机采集瑕疵图片并对其进行标注，制作瑕疵数据集，利用YOLOv5s模型进行目标检测。使用ImgAug数据增强技术扩充原始数据集的数量，并在训练过程中结合Mosaic数据增强方法解决数据样本少的问题。在YOLOv5的基础上，使用轻量化的GhostNet模块替换主干网络和颈部的卷积模块，有效减少模型的参数量和计算量，通过改进激活函数减少网络复杂度并加快计算速度，以满足实时性需求。在主干网络中加入新型注意力机制Polarized Self-Attention模块，增强网络对于小目标瑕疵的特征提取能力。实验结果表明，相比YOLOv5，GPC-YOLOv5算法的参数量和计算量分别减少25.4%和28.5%，总体mAP达到89.2%，能够有效提高检测精度并加快检测速度。

关键词: 皮革瑕疵, 轻量型YOLOv5算法, 注意力机制, 深度学习, 目标检测

Abstract:

In the industrial leather production industry, leather defect detection is an important part of quality control. Aiming at the problems of high computational complexity, difficulties in small target defect detection, and few data samples, this study proposes an improved lightweight leather defect detection algorithm based on attention mechanism, GPC-YOLOv5. Defect pictures are collected using an industrial camera; whereby the images are annotated to create a defect dataset, and use the YOLOv5s model for object detection.The number of original datasets are expanded to solve the problem of few data samples using the Mosaic method combined with ImgAug data augmentation in the training process.The lightweight GhostNet module replaces the convolution module of the backbone network and neck in YOLOv5, thereby effectively reducing the number of parameters and computation load of the model.The network complexity reduces and calculations accelerate because of improving the activation function and real-time requirements are satisfied. The feature extraction ability of the network for small target defects is enhanced by adding a new attention mechanism, Polarized Self-Attention module, to the backbone network.The experimental results show that compared to YOLOv5, the GPC-YOLOv5 algorithm reduces the number of parameters and computation load by 25.4% and 28.5%, respectively, with an overall mean Average Precision(mAP) of 89.2%, effectively improving detection accuracy and accelerating detection speed.

Key words: leather defect, lightweight YOLOv5 algorithm, attention mechanism, deep learning, object detection

刘俊豪, 王美林, 谢兴, 宋烨兴, 许莉花. 基于改进YOLOv5的皮革瑕疵检测算法[J]. 计算机工程, 2023, 49(8): 240-249.

Junhao LIU, Meilin WANG, Xing XIE, Yexing SONG, Lihua XU. Leather Defect Detection Algorithm Based on Improved YOLOv5[J]. Computer Engineering, 2023, 49(8): 240-249.

https://www.ecice06.com/CN/Y2023/V49/I8/240

图/表 18

图1 YOLOv5模型结构

Fig.1 Structure of YOLOv5 model

图2 使用常规卷积得到的特征图

Fig.2 Feature maps obtained using conventional convolution

图3 Ghost卷积模块结构

Fig.3 Structure of Ghost convolution module

图4 使用Ghost模块得到的特征图

Fig.4 Feature maps obtained using Ghost module

图5 Polarized Self-Attention模块结构

Fig.5 Structure of Polarized Self-Attention module

图6 DIoU的计算示意图

Fig.6 Schematic diagram of DIoU calculation

图7 改进边界框损失函数的检测效果对比

Fig.7 Detection effect comparison of improved bounding box loss functions

图8 皮革瑕疵数据集标注样例

Fig.8 Annotation example of leather defect dataset

表1 在数据集中各类瑕疵的统计结果

Table 1 Statistical results of each type of defect on the dataset

数据集	图片数/张	气泡状/个	孔洞状/个	条纹状/个	Label总数/个
训练集	636	1 252	807	484	2 543
测试集	158	292	178	125	595
总计	794	1 544	985	609	3 138

图9 ImgAug数据增强方法

Fig.9 ImgAug data enhancement methods

图10 Mosaic数据增强示意图

Fig.10 Schematic diagram of Mosaic data enhancement

表2 替换GhostNet前后模型的参数信息

Table 2 Parameter information of models before and after replacing GhostNet

模型	FLOPs/10⁹	参数量/10⁶	模型大小/MB
YOLOv5	15.8	7.02	13.8
Ghost-YOLOv5	8.1	3.68	7.5

表3 不同模块的实验结果对比

Table 3 Experiment results comparison among different modules

模块	FLOPs/10⁹	参数量/10⁶	召回率	F1值	mAP/%
Ghost-YOLOv5	8.1	3.68	0.786	0.834	86.0
Ghost-YOLOv5+CBAM	10.0	4.50	0.782	0.838	86.6
Ghost-YOLOv5+SENet	10.0	4.54	0.804	0.850	87.7
Ghost-YOLOv5+ParNetAttention	18.6	8.29	0.811	0.852	87.8
Ghost-YOLOv5+CotAttention	17.3	7.51	0.793	0.839	86.6
Ghost-YOLOv5+ECAttention	9.9	4.45	0.815	0.858	88.3
Ghost-YOLOv5+Polarized Self-Attention	11.3	5.16	0.809	0.859	88.9
Ghost-YOLOv5+EIoU	8.1	3.68	0.754	0.811	84.4
Ghost-YOLOv5+AlaphIoU (α=3)	8.1	3.68	0.755	0.821	83.4
Ghost-YOLOv5+CDIoU	8.1	3.68	0.788	0.839	86.5
Ghost-YOLOv5+Polarized Self-Attention+EIoU	11.3	5.16	0.800	0.845	87.8
Ghost-YOLOv5+Polarized Self-Attention+AlaphIoU	11.3	5.16	0.787	0.844	86.4
Ghost-YOLOv5+ECAttention+CDIoU	9.9	4.45	0.817	0.860	88.7
Ghost-YOLOv5+Polarized Self-Attention+CDIoU	11.3	5.16	0.819	0.867	89.2

表4 消融实验结果

Table 4 Ablation experimental results

Ghost-YOLOv5	Polarized Self-Attention	CDIoU	召回率	F1值		AP/%		mAP/%
Ghost-YOLOv5	Polarized Self-Attention	CDIoU	召回率	F1值	气泡状	孔洞状	条纹状	mAP/%
√	×	×	0.786	0.834	90.7	83.8	83.4	86.0
√	√	×	0.809	0.859	93.6	86.9	86.1	88.9
√	×	√	0.788	0.839	92.2	82.9	84.5	86.5
√	√	√	0.819	0.867	93.8	87.1	86.6	89.2

图11 改进前后数据集的指标对比

Fig.11 Comparison of indicators in the dataset before and after improvement

表5 不同算法的评价指标对比

Table 5 Comparison of evaluation indicators for different algorithms

算法	参数量/10⁶	FPS	模型大小/MB	mAP/%
Faster-RCNN	41.13	23.8	315.1	46.4
Cascade-RCNN	68.93	15.8	527.6	51.4
Libra-RCNN	41.40	22.7	317.1	47.1
YOLOF	42.11	38.4	322.8	47.2
RetinaNet	36.15	19.6	277.0	62.6
VFNet	32.49	21.3	249.1	74.1
YOLOv3	61.53	32.2	469.9	80.8
YOLOv3-MobileNetV2	3.67	52.6	26.8	71.2
YOLOX	8.94	35.7	102.9	85.7
本文算法	5.16	83.3	10.3	89.2

图12 不同算法的检测效果对比

Fig.12 Detection effects comparison among different algorithms

图13 不同方法的ROC曲线及对应的AUC值

Fig.13 ROC curves and the corresponding AUC values among different methods

参考文献 31

1	FU D H, DAI L, DAI J H. Automatic detection and localization of surface defects for whole piece of ultrahigh-definition leather images[C]//Proceedings of the 4th International Conference on Computer and Communication Systems. Washington D. C., USA: IEEE Press, 2019: 229-232.
2	MOHAMMED K M C, KUMAR S S, PRASAD G. Optimized fuzzy C-means clustering methods for defect detection on leather surface. Journal of Scientific & Industrial Research, 2020, 79 (9): 883- 836.
3	KUMAR M P, ASHOK D. A multi-level colour thresholding based segmentation approach for improved identification of the defective region in leather surfaces. Engineering Journal, 2020, 24 (2): 101- 108. doi: 10.4186/ej.2020.24.2.101
4	KAYNAR O, IŞIK Y E, GÖRMEZ Y, et al. Fabric defect detection with LBP-GLMC[C]//Proceedings of International Artificial Intelligence and Data Processing Symposium. Washington D. C., USA: IEEE Press, 2017: 1-5.
5	ZHANG N, XIANG J, WANG L, et al. Image retrieval of wool fabric. Part Ⅰ: based on low-level texture features. Textile Research Journal, 2019, 89 (19/20): 4195- 4207.
6	LI N, ZHAO J Y, JIANG P. Fabric defects detection via visual attention mechanism[C]//Proceedings of Chinese Automation Congress. Washington D. C., USA: IEEE Press, 2018: 2956-2960.
7	SADAGHIYANFAM S. Using gray-level-co-occurrence matrix and wavelet transform for textural fabric defect detection: a comparison study[C]//Proceedings of Electric Electronics, Computer Science, Biomedical Engineerings' Meeting. Washington D. C., USA: IEEE Press, 2018: 1-5.
8	MOHAMMED K M C, SRINIVAS KUMAR S, PRASAD G. Defective texture classification using optimized neural network structure. Pattern Recognition Letters, 2020, 135, 228- 236. doi: 10.1016/j.patrec.2020.04.017
9	PEREIRA R F, MEDEIROS C M S, FILHO P P R. Goat leather quality classification using computer vision and machine learning[C]//Proceedings of International Joint Conference on Neural Networks. Washington D. C., USA: IEEE Press, 2018: 1-8.
10	JAWAHAR M, CHANDRA BABU N K, VANI K, et al. Vision based inspection system for leather surface defect detection using fast convergence particle swarm optimization ensemble classifier approach. Multimedia Tools and Applications, 2021, 80 (3): 4203- 4235. doi: 10.1007/s11042-020-09727-3
11	AI J Q, MAO Y X, LUO Q W, et al. SAR target classification using the multikernel-size feature fusion-based convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60, 1- 10.
12	LIONG S T, ZHENG D N, HUANG Y C, et al. Leather defect classification and segmentation using deep learning architecture. International Journal of Computer Integrated Manufacturing, 2020, 33 (10/11): 1105- 1117.
13	LIU J, ZHANG B G, LI L. Defect detection of fabrics with generative adversarial network based flaws modeling[C]//Proceedings of Chinese Automation Congress. Washington D. C., USA: IEEE Press, 2021: 3334-3338.
14	CHEN S Y, CHENG Y C, YANG W L, et al. Surface defect detection of wet-blue leather using hyperspectral imaging. IEEE Access, 2021, 9, 127685- 127702. doi: 10.1109/ACCESS.2021.3112133
15	赵亚男, 吴黎明, 陈琦. 基于多尺度融合SSD的小目标检测算法. 计算机工程, 2020, 46 (1): 247- 254. URL
	ZHAO Y N, WU L M, CHEN Q. Small object detection algorithm based on multi-scale fusion SSD. Computer Engineering, 2020, 46 (1): 247- 254. URL
16	程珍. 基于深度卷积神经网络的织物瑕疵检测算法研究[D]. 上海: 东华大学, 2021.
	CHENG Z. Research on fabric defect detection algorithm based on deep convolutional neural network[D]. Shanghai: Donghua University, 2021. (in Chinese)
17	杨毅, 桑庆兵. 多尺度特征自适应融合的轻量化织物瑕疵检测. 计算机工程, 2022, 48 (12): 288- 295. URL
	YANG Y, SANG Q B. Lightweight-fabric defect detection based on adaptive fusion of multiscale features. Computer Engineering, 2022, 48 (12): 288- 295. URL
18	彭成, 张乔虹, 唐朝晖, 等. 基于YOLOv5增强模型的口罩佩戴检测方法研究. 计算机工程, 2022, 48 (4): 39- 49. URL
	PENG C, ZHANG Q H, TANG Z H, et al. Research on mask wearing detection method based on YOLOv5 enhancement model. Computer Engineering, 2022, 48 (4): 39- 49. URL
19	HAN K, WANG Y H, TIAN Q, et al. GhostNet: more features from cheap operations[C]//Proceedings of Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2020: 1577-1586.
20	LIU H J, LIU F Q, FAN X Y, et al. Polarized Self-Attention: towards high-quality pixel-wise regression[EB/OL]. [2022-03-25]. https://arxiv.org/abs/2107.00782.
21	CHEN D, MIAO D Q. Control distance IoU and control distance IoU loss function for better bounding box regression[EB/OL]. [2022-03-25]. https://arxiv.org/abs/2103.11696.
22	AI J Q, WANG F F, MAO Y X, et al. A fine PolSAR terrain classification algorithm using the texture feature fusion-based improved convolutional autoencoder. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60, 1- 14.
23	徐先峰, 赵万福, 邹浩泉, 等. 基于MobileNet-SSD的安全帽佩戴检测算法. 计算机工程, 2021, 47 (10): 298-305, 313 URL
	XU X F, ZHAO W F, ZOU H Q, et al. Detection algorithm of safety helmet wear based on MobileNet-SSD. Computer Engineering, 2021, 47 (10): 298-305, 313 URL
24	GOYAL A, BOCHKOVSKIY A, DENG J, et al. Non-deep networks[EB/OL]. [2022-03-25]. https://arxiv.org/abs/2110.07641.
25	LI Y H, YAO T, PAN Y W, et al. Contextual transformer networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45 (2): 1489- 1500.
26	WANG Q L, WU B G, ZHU P F, et al. ECA-Net: efficient channel attention for deep convolutional neural networks[C]//Proceedings of Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2020: 1-10.
27	ZHANG Y F, REN W Q, ZHANG Z, et al. Focal and efficient IoU loss for accurate bounding box regression[EB/OL]. [2022-03-25]. https://arxiv.org/abs/2101.08158.
28	HE J B, ERFANI S, MA X J, et al. Alpha-IoU: a family of power intersection over union losses for bounding box regression[EB/OL]. [2022-03-25]. https://arxiv.org/abs/2110.13675.
29	CHEN Q, WANG Y M, YANG T, et al. You only look one-level feature[C]//Proceedings of Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2021: 13034-13043.
30	ZHANG H Y, WANG Y, DAYOUB F, et al. VarifocalNet: an IoU-aware dense object detector[C]//Proceedings of Conference on Computer Vision and Pattern Recognition. Washington D. C., USA. IEEE Press, 2021: 8510-8519.
31	GE Z, LIU S T, WANG F, et al. YOLOX: exceeding YOLO series in 2021[EB/OL]. [2022-03-25]. https://arxiv.org/abs/2107.08430.

[1]	王炼红, 林飞鹏, 李潇瑶, 谌桂枝, 周莉. 融入课程知识图谱的KMAKT预测[J]. 计算机工程, 2024, 50(7): 23-31.
[2]	钱来, 赵卫伟. 基于对比学习和注意力机制的文本分类方法[J]. 计算机工程, 2024, 50(7): 104-111.
[3]	刘建敏, 林晖, 汪晓丁. 基于图注意力机制的无地图场景轨迹预测方法[J]. 计算机工程, 2024, 50(7): 144-153.
[4]	牛瑞婷, 严天峰, 高锐, 王映植. 低信噪比下基于深度学习TCNN-MobileNet的调制识别[J]. 计算机工程, 2024, 50(7): 204-215.
[5]	屠乃威, 焦猛, 阎馨. 复杂环境下输电线路鸟巢目标图像检测模型[J]. 计算机工程, 2024, 50(7): 216-226.
[6]	肖慈, 徐杨, 张永丹, 冯明文, 黄易仟. 结合注意力和低光增强的夜间语义分割[J]. 计算机工程, 2024, 50(7): 271-281.
[7]	张诗婧, 莫绪涛, 赵行, 董杨林. 基于球面折反射成像和YOLOv7的内螺纹缺陷检测[J]. 计算机工程, 2024, 50(7): 282-292.
[8]	张锡英, 孙守东, 于海浩, 边继龙. 基于空间传播的多视图三维重建[J]. 计算机工程, 2024, 50(7): 293-302.
[9]	贵向泉, 刘世清, 李立, 秦庆松, 李唐艳. 基于改进YOLOv8的景区行人检测算法[J]. 计算机工程, 2024, 50(7): 342-351.
[10]	徐明亮, 李芳媛, 马浩然, 何飞. 大规模神经记录的峰电位聚类算法(特邀)[J]. 计算机工程, 2024, 50(6): 1-34.
[11]	魏琢艺, 罗迈, 李文兵, 曾远松, 余伟江, 杨跃东. 基于多源域适应的单细胞智能分类[J]. 计算机工程, 2024, 50(6): 48-55.
[12]	李子杰, 周菊香, 韩晓瑜, 甘健侯, 鹿泽光, 王俊. 序列特征与学习过程融合的知识追踪模型[J]. 计算机工程, 2024, 50(6): 77-85.
[13]	程腾腾, 姚春龙, 于晓强, 李旭, 王庆丰. 基于多头注意力机制融合常识知识的共情对话生成[J]. 计算机工程, 2024, 50(6): 94-101.
[14]	更藏措毛, 黄鹤鸣, 杨毅杰. 融合多尺度特征与上下文信息的语音增强方法[J]. 计算机工程, 2024, 50(6): 138-147.
[15]	李永飞, 李铭洋, 常鑫, 曹可欣. 基于可解释性深度学习的物联网水质监测数据异常检测[J]. 计算机工程, 2024, 50(6): 179-187.

选择文件类型/文献管理软件名称

选择包含的内容