基于PP-PicoDet-XS的改进铝型材表面缺陷检测算法

doi:10.12068/j.issn.1005-3026.2024.11.005

摘要/Abstract

摘要：

铝型材在生产加工过程中会产生特征不明显和尺度大小不一等多类型的表面缺陷，针对现有人工抽检方法准确率低、实时性差、主观性强等问题，提出一种基于PP-PicoDet-XS的改进铝型材表面缺陷检测算法.改进的算法在主干网络中嵌入无参注意力SimAM，增强对深层有效特征的提取能力；使用SIoU（Scylla intersection over union）损失函数对训练过程进行优化，提高预测框的定位能力；采用量化蒸馏策略对模型进行压缩，提高推理速度.结果表明，改进的算法平均精度均值在交并比（intersection over union，IoU）阈值为0.5时达到了98.93%，在IoU阈值0.5~0.95范围内达到了57.60%，较未压缩的原始模型分别提高了1.73%和4.13%.将该算法部署到骁龙865移动端平台上进行推理，推理速度可达116.82帧/s，较未压缩的原始模型提高了47帧/s.

关键词: 铝型材, 缺陷检测, SimAM, 损失函数, 量化, 蒸馏

Abstract:

During the production and processing of aluminum profiles， multiple types of surface defects such as unclear features and varying scales may generate. In response to the problems of low accuracy， poor real?time performance， and strong subjectivity in existing manual sampling method， an improved surface defects detection algorithm is proposed for aluminum profiles based on PP-PicoDet-XS.The SimAM attention was embedded in the backbone to enhance the ability of extracting deep effective features. The SIoU（Scylla intersection over union） loss function is used to optimize the training process to improve the positioning ability of the prediction boxes. The quantization and distillation were used to compress the model to improve the inference speed. The results show that the improved algorithm achieves a mean average precision of 98.93% at intersection over union（IoU） threshold of 0.5， and 57.60% across IoU thresholds ranging from 0.5 to 0.95， which is 1.73% and 4.13% higher than the uncompressed original model. Deploying this algorithm on the Snapdragon 865 mobile platform for inference， the inference speed can reach 116.82 frames per second， which is 47 frames per second higher than the uncompressed original model.

Key words: aluminum profiles, defects detection, SimAM, loss function, quantization, distillation

中图分类号:

TP 391.4

马淑华, 李立振, 秦汉民, 沙晓鹏. 基于PP-PicoDet-XS的改进铝型材表面缺陷检测算法[J]. 东北大学学报（自然科学版）, 2024, 45(11): 1557-1564.

Shu-hua MA, Li-zhen LI, Han-min QIN, Xiao-peng SHA. Improved Surface Defects Detection Algorithm for Aluminum Profiles Based on PP-PicoDet-XS[J]. Journal of Northeastern University(Natural Science), 2024, 45(11): 1557-1564.

图/表 11

参考文献 24

1	Wei X K， Yang Z M， Liu Y X，et al.Railway track fastener defect detection based on image processing and deep learning techniques：a comparative study［J］.Engineering Applications of Artificial Intelligence，2019，80：66-81.
2	Jian C X， Gao J， Ao Y H.Automatic surface defect detection for mobile phone screen glass based on machine vision［J］.Applied Soft Computing，2016，52（3）：348-358.
3	赵翔宇，周亚同，何峰，等.工业干扰环境下基于模板匹配的印刷品缺陷检测［J］.包装工程，2017，38（11）：187-192.
	Zhao Xiang‑yu， Zhou Ya‑tong， He Feng，et al.Printing defects detection based on template matching under disturbing industrial environment［J］.Packaging Engineering，2017，38（11）：187-192.
4	李永敬，朱萍玉，孙孝鹏，等.基于形状模板匹配的冲压件外形缺陷检测算法研究［J］.广州大学学报（自然科学版），2017，16（5）：62-66.
	Li Yong‑jing， Zhu Ping‑yu， Sun Xiao‑peng，et al.Shape defect detection algorithm of stamping parts based on shape template matching［J］.Journal of Guangzhou University （Natural Science Edition），2017，16（5）：62-66.
5	孙光民，刘鹏，李子博.基于图像处理的带钢表面缺陷检测改进算法的研究［J］.软件工程，2018，21（4）：5-8.
	Sun Guang‑min， Liu Peng， Li Zi‑bo.Research on the detection algorithm of strip steel surface defects based on image processing［J］.Software Engineering，2018，21（4）：5-8.
6	Girshick R， Donahue J， Darrell T，et al.Rich feature hierarchies for accurate object detection and semantic segmentation［C］//IEEE Conference on Computer Vision and Pattern Recognition.Columbus：IEEE，2014：580-587.
7	Girshick R.Fast R-CNN［C］//IEEE International Conference on Computer Vision.Santiago：IEEE，2015：1440-1448.
8	Ren S Q， He K M， 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.
9	He K M， Gkioxari G， Dollár P，et al.Mask R-CNN［C］//2017 IEEE International Conference on Computer Vision （ICCV）.Venice，2017：2980-2988.
10	Liu W， Anguelov D， Erhan D，et al.SSD：single shot multibox detector［C］//European Conference on Computer Vision.Berlin：Springer，2016：21-37.
11	Redmon J， Divvala S， Girshick R，et al.You only look once：unified，real‑time object detection［C］//IEEE Conference on Computer Vision and Pattern Recognition.Las Vegas：IEEE，2016：779-788.
12	Howard A G， Zhu M L， Chen B，et al.MobileNets：efficient convolutional neural networks for mobile vision applications［EB/OL］.（2017-04-17）［2023-04-03］. .
13	Zhang X Y， Zhou X Y， Lin M X，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	Han K， Wang Y H， Tian Q，et al.GhostNet：more features from cheap operations［C］//IEEE Conference on Computer Vision and Pattern Recognition.Seattle：IEEE，2020：1577-1586.
15	王淑青，鲁濠，鲁东林，等.基于轻量化人工神经网络的PCB板缺陷检测［J］.仪表技术与传感器，2022（5）：98-104.
	Wang Shu‑qing， Lu Hao， Lu Dong‑lin，et al.PCB board defect detection based on lightweight artificial neural network［J］.Instrument Technique and Sensor，2022（5）：98-104.
16	Ge Z， Liu S， Wang F，et al.YOLOX：exceeding YOLO series in 2021［EB/OL］.（2021-08-06）［2023-04-08］..
17	Wang C Y， Bochkovskiy A， Liao H Y M.YOLOv7：trainable bag‑of‑freebies sets new state‑of‑the‑art for real‑time object detectors［C］//IEEE Conference on Computer Vision and Pattern Recognition.Vancouver：IEEE，2023：7464-7475.
18	Yu G， Chang Q， Lyu W，et al.PP-PicoDet：a better real‑time object detector on mobile devices［EB/OL］.（2021-11-01）［2023-04-15］..
19	Cui C， Gao T Q， Wei S Y，et al.PP-LCNet：a lightweight CPU convolutional neural network［EB/OL］.（2021-09-17）［2023-04-15］..
20	Rezatofighi H， Tsoi N， Gwak J，et al ．Generalized intersection over union：a metric and a loss for bounding box regression［C］//IEEE Conference on Computer Vision and Pattern Recognition．New York：IEEE，2019：658-666．
21	Yang L， Zhang R Y， Li L，et al.SimAM：a simple，parameter‑free attention module for convolutional neural networks［C］//International Conference on Machine Learning.［n.l］：PMLR，2021：11863-11874.
22	Gevorgyan Z.SIoU loss：more powerful learning for bounding box regression［EB/OL］.（2022-05-15）［2023-04-23］..
23	Jacob B， Kligys S， Chen B，et al.Quantization and training of neural networks for efficient integer‑arithmetic‑only inference［C］//IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City：IEEE，2018：2704-2713.
24	Hinton G， Vinyals O， Dean J.Distilling the knowledge in a neural network［J］.Computer Science，2015，14（7）：38-39.

主干网络组成	算子	卷积核尺寸	步长	SE
卷积块	卷积	3×3	2	—
DSC模块_1	DSC	3×3	1	—
DSC模块_2	DSC	3×3	2	—
DSC模块_2	DSC	3×3	1	—
DSC模块_3	DSC	3×3	2	—
DSC模块_3	DSC	3×3	1	—
DSC模块_4	DSC	3×3	2	—
DSC模块_4	DSC×5	5×5	1	—
DSC模块_5	DSC	5×5	2	√
DSC模块_5	DSC	5×5	1	√

主干网络组成	算子	卷积核尺寸	步长	SE
卷积块	卷积	3×3	2	—
DSC模块_1	DSC	3×3	1	—
DSC模块_2	DSC	3×3	2	—
DSC模块_2	DSC	3×3	1	—
DSC模块_3	DSC	3×3	2	—
DSC模块_3	DSC	3×3	1	—
DSC模块_4	DSC	3×3	2	—
DSC模块_4	DSC×5	5×5	1	—
DSC模块_5	DSC	5×5	2	√
DSC模块_5	DSC	5×5	1	√

网络模型	SimAM 位置	mAP@0.5∶0.95	mAP@0.5
网络模型	SimAM 位置	%	%
PP-PicoDet- XS	1 0 0 0 0	54.51	98.47
	0 1 0 0 0	54.12	98.02
	0 0 1 0 0	53.48	97.51
	0 0 0 1 0	53.55	97.22
	0 0 0 0 1	55.56	98.81
	1 1 1 1 1	53.66	97.94

网络模型	SimAM 位置	mAP@0.5∶0.95	mAP@0.5
网络模型	SimAM 位置	%	%
PP-PicoDet- XS	1 0 0 0 0	54.51	98.47
	0 1 0 0 0	54.12	98.02
	0 0 1 0 0	53.48	97.51
	0 0 0 1 0	53.55	97.22
	0 0 0 0 1	55.56	98.81
	1 1 1 1 1	53.66	97.94

网络模型	IoU 损失函数	mAP@0.5∶0.95	mAP@0.5
网络模型	IoU 损失函数	%	%
PP-PicoDet-XS	GIoU	53.47	97.20
	DIoU	54.16	97.73
	CIoU	54.06	97.89
	SIoU	56.22	98.72