基于动态视觉传感器的航发叶片缺陷检测

doi:10.11996/JG.j.2095-302X.2026010120

摘要/Abstract

摘要：

航空发动机叶片作为发动机核心零部件，其表面微小缺陷可能导致严重安全事故，传统视觉检测技术受限于运动模糊、动态范围低及背景冗余等问题。针对上述挑战，提出一种基于动态视觉传感器(DVS)的航发叶片缺陷检测方法。动态视觉传感器数据格式为异步事件流，故也被称作事件相机，具有动态范围大、高帧率和微小目标捕捉能力强等优势。首先搭建基于DVS的缺陷检测平台，探索总结了其成像特点及优势。在此基础上，构建首个基于DVS的航发叶片缺陷检测数据集(EDD-AB)，涵盖划痕、点痕、边缘损伤3类缺陷近6 000张图像，精细标注近1.2万个目标标签，数据集已开源(链接： https://github.com/NiBieZhouMei5520/EDD-AB.git)。进一步提出基于异步事件流帧聚合的多尺度缺陷检测算法(AEAF-ABDD)：通过固定时间窗的帧聚合技术实现事件流可视化；构建多分辨率自适应特征金字塔网络(MRAFPN)增强多尺度缺陷特征提取能力；引入轻量级SimAM注意力机制强化关键区域聚焦；融合星形卷积模块(StarNet)提升高维非线性特征映射效率，实现复杂曲面工件多尺度缺陷的精准检测。实验表明，AEAF-ABDD在EDD-AB数据集上的平均精度均值(mAP)达97.7%，检测速度达105帧/秒，显著优于主流算法，为高反光曲面工件的自动化质检提供了高效解决方案，推动了DVS在工业检测领域的应用。

关键词: 动态视觉传感器, 航空发动机叶片, 缺陷检测, 异步事件流, 多尺度特征融合

Abstract:

Aeroengine blades are core components of engines; tiny surface defects can lead to serious safety accidents. Traditional vision detection technology is limited by motion blur, low dynamic range, background redundancy, and so forth. To address these challenges, a method of aeroengine blade defect detection based on Dynamic Vision Sensor (DVS) was proposed. Dynamic vision sensor produced data in an asynchronous event-stream format, and were therefore referred to as event camera, which exhibited the advantages of large dynamic range, high frame rate, and strong ability to capture small targets. Firstly, a defect detection platform based on DVS was built, and its imaging characteristics and advantages were explored. On this basis, the first Event-based Defect Detection Dataset of Aeroengine Blade (EDD-AB) dataset based on DVS was constructed, covering nearly 6 000 images of scratches, point marks and edge damage, with approximately 12 000 finely annotated target labels. The dataset was released as open source (link: https://github. com/NiBieZhouMei5520/EDD-AB.git). Furthermore, a multi-scale defect-detection algorithm based on asynchronous event-stream frame aggregation (AEAF-ABDD) was proposed: event streams were visualized through frame aggregation technology using a fixed time window; a Multi-Resolution Adaptive Feature Pyramid Network (MRAFPN) was developed to enhance multi-scale defect feature extraction capability; a lightweight SimAM attention mechanism was incorporated to strengthen focus on key regions; a star-convolution module (StarNet) was fused to improve the efficiency of high-dimensional nonlinear feature mapping, enabling accurate detection of multi-scale defects on complex curved workpieces. Experiments demonstrated that AEAF-ABDD achieved a mean Average Precision (mAP) of 97.7% on the EDD-AB dataset and a detection speed of 105 frames per second, substantially outperforming mainstream algorithms. An efficient solution for automated quality inspection of highly reflective curved workpieces was thereby provided, promoting the application of DVS in the field of industrial inspection.

Key words: dynamic vision sensor, aeroengine blades, defect detection, asynchronous event stream, multi-scale feature

中图分类号:

张行顺, 陈海永. 基于动态视觉传感器的航发叶片缺陷检测[J]. 图学学报, 2026, 47(1): 120-130.

ZHANG Xingshun, CHEN Haiyong. Defect detection of aero-engine blades based on dynamic vision sensors[J]. Journal of Graphics, 2026, 47(1): 120-130.

图/表 15

图1 事件产生原理图

Fig. 1 Schematic diagram of the event generation principle

图2 缺陷成像示意图((a) 工件位置未发生变化时的光路；(b) 工件与DVS发生相对位移时的光路)

Fig. 2 Schematic diagram of defect imaging ((a) Optical path with a stationary workpiece；(b) The optical path during relative displacement between the workpiece and the DVS)

图3 实验平台

Fig. 3 Experimental platform

图4 弱小缺陷成像对比(红框中为表面划痕缺陷)((a) DVS成像效果；(b) 传统相机成像效果)

Fig. 4 Comparison of imaging of weak defects (surface scratch defects in red box) ((a) DVS imaging effect; (b) Traditional camera imaging effect)

图5 高动态范围成像对比(红框中为表面划痕缺陷与点痕缺陷) ((a) DVS成像效果；(b) 传统相机成像效果)

Fig. 5 High dynamic range imaging comparison (surface scratch defects and point scratch defects in the red box) ((a) DVS imaging effect; (b) Traditional camera imaging effect)

图6 不同类型缺陷的事件流可视化图像(红色方框中为划痕，红色圆圈中为点痕，红色三角形中为边缘损伤)

Fig. 6 Visualization images of event flows for different types of defects (the scratches are in the red boxes, the dot marks are in the red circles, and the edge damages are in the red triangles)

图7 数据集 EDD-AB 特性统计((a) 划痕损伤目标边界框宽高比的分布；(b) 点痕损伤目标边界框宽高比的分布；(c) 边缘损伤目标边界框宽高比的分布；(d) 所有缺陷目标边界框的面积分布)

Fig. 7 EDD-AB Characteristics Statistics of Dataset ((a) Area distribution of bounding boxes for seratches; (b) Area distribution of bounding boxes for point marks; (c) Area distribution of bounding boxes for edge damage; (d) Area distribution of bounding boxes for all defect targets)

图8 小目标检测头

Fig. 8 Small target detection head

图9 SimAM注意力结构示意图

Fig. 9 Schematic diagram of the SimAM attention structure

图10 星形卷积操作示意图(*为星形运算)

Fig. 10 Schematic diagram of star-shaped convolution operation(* is element-wise multiplication)

表1 实验环境配置

Table 1 Experimental environment configuration

配置	参数
操作环境	Windows11
深度学习框架	Pytorch1.12
CUDA	11.3
Python版本	Python-3.12
CPU	Intel(R) Core(TM) i7-14700HX 2.10 GHz
GPU	NVIDIA GeForce RTX 4070Laptop 8 GB

表2 实验参数设置

Table 2 Experimental parameter settings

超参数	参数值	超参数	参数值
Images size	640×640	Optimize	SGD
Epochs	300	Momentum	0.973
Batch size	32	Learning rate	0.01

表3 与其他算法在EDD-AB数据集上的对比实验

Table 3 Comparative experiments with other algorithms on the EDD-AB dataset

Method	P/%	R/%	Parms/10⁶	GFLOPs	FPS	mAP/%
Method	P/%	R/%	Parms/10⁶	GFLOPs	FPS	综合	划痕	点痕	边缘损伤
Faster-RCNN	76.6	77.6	165.00	199.0	66	80.2	87.7	66.2	86.6
SSD	63.0	65.3	24.50	87.9	72	71.0	77.6	55.3	79.6
DETR	77.5	78.0	9.49	16.8	88	87.0	91.3	77.5	91.8
Yolov10	89.6	87.5	2.59	6.4	96	92.8	97.1	85.7	96.0
Yolov12	85.3	88.7	2.52	6.0	93	91.8	94.8	89.1	92.2
EMS-YOLO^[18]	82.3	84.5	14.40	6.8	90	88.8	90.2	83.3	91.9
TIFF-EDD^[13]	90.6	91.1	3.06	28.4	82	94.1	95.8	89.9	96.5
AEAF-ABDD	93.2	93.9	3.00	12.8	105	97.7	98.6	95.9	97.9

图11 检测热力图对比

Fig. 11 Comparison of hatemap detection ((a) Original image; (b) DETR; (c) Yolov10; (d) TIFF-EDD; (e) AEAF-ABDD)

表4 在EDD-AB数据集上的消融实验

Table 4 Ablation experiments on the EDD-AB dataset

Base	MSAFPN	SimAM	StarNet	P/%	R/%	Parms/10⁶	GFLOPs	FPS	mAP/%
Base	MSAFPN	SimAM	StarNet	P/%	R/%	Parms/10⁶	GFLOPs	FPS	综合	划痕	点痕	边缘损伤
√				87.1	89.6	3.00	8.2	99	92.9	96.4	87.8	94.7
√	√			92.6	93.0	2.93	12.4	100	94.0	96.4	87.6	96.5
√		√		89.6	90.8	3.01	8.2	98	93.6	95.4	90.9	97.0
√			√	88.6	90.5	3.08	8.5	89	94.1	96.9	89.2	95.8
√	√	√		91.9	92.3	2.93	12.4	101	95.5	97.6	93.2	96.5
√		√	√	90.3	90.8	3.08	8.5	96	94.2	96.3	88.6	96.6
√	√		√	91.6	92.5	3.00	12.8	93	95.2	95.8	89.7	97.3
√	√	√	√	93.2	93.9	3.00	12.8	105	97.7	98.6	95.9	97.9

参考文献 26

[1]	陶显, 侯伟, 徐德. 基于深度学习的表面缺陷检测方法综述[J]. 自动化学报, 2021, 47(5): 1017-1034.
	TAO X, HOU W, XU D. A survey of surface defect detection methods based on deep learning[J]. Acta Automatica Sinica, 2021, 47(5): 1017-1034 (in Chinese).
[2]	CHEN Y Q, PAN J W, LEI J Y, et al. EEE-Net: efficient edge enhanced network for surface defect detection of glass[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 5029013.
[3]	JIANG W B, LIU M, PENG Y N, et al. HDCB-Net: a neural network with the hybrid dilated convolution for pixel-level crack detection on concrete bridges[J]. IEEE Transactions on Industrial Informatics, 2021, 17(8): 5485-5494. DOI URL
[4]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot MultiBox detector[C]// The 14th European Conference on Computer Vision. Cham: Springer, 2016: 21-37.
[5]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2016: 779-788.
[6]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]// 2014 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2014: 580-587.
[7]	ZHENG Y J, ZHENG L X, YU Z F, et al. High-speed image reconstruction through short-term plasticity for spiking cameras[C]// 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2021: 6354-6363.
[8]	XU X Y, SUN D Q, PAN J S, et al. Learning to super-resolve blurry face and text images[C]// 2017 IEEE International Conference on Computer Vision. New York: IEEE Press, 2017: 251-260.
[9]	李家宁, 田永鸿. 神经形态视觉传感器的研究进展及应用综述[J]. 计算机学报, 2021, 44(6): 1258-1286.
	LI J N, TIAN Y H. Recent advances in neuromorphic vision sensors: a survey[J]. Chinese Journal of Computers, 2021, 44(6): 1258-1286 (in Chinese).
[10]	LICHTSTEINER P, POSCH C, DELBRUCK T. A 128 × 128 120db 30mw asynchronous vision sensor that responds to relative intensity change[C]// 2006 IEEE International Solid State Circuits Conference-Digest of Technical Papers. New York: IEEE Press, 2006: 2060-2069.
[11]	CHEN S S, GUO M H. Live demonstration: CeleX-V: a 1M pixel multi-mode event-based sensor[C]// 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. New York: IEEE Press, 2019: 1682-1683.
[12]	BRANDLI C, MULLER L, DELBRUCK T. Real-time, high-speed video decompression using a frame- and event-based DAVIS sensor[C]// 2014 IEEE International Symposium on Circuits and Systems. New York: IEEE Press, 2014: 686-689.
[13]	马居坡, 陈周熠, 吴金建. 基于动态视觉传感器的铝基盘片表面缺陷检测[J]. 自动化学报, 2024, 50(12): 2407-2419.
	MA J P, CHEN Z Y, WU J J. Dynamic vision sensor based defect detection for the surface of aluminum disk[J]. Acta Automatica Sinica, 2024, 50(12): 2407-2419 (in Chinese).
[14]	吕和平, 洪流. 基于事件相机的铝基片划痕检测研究[J]. 数字自造科学, 2024, 22(2): 140-145.
	LV H P, HONG L. Research on scratch detection of aluminum substrate based on event camera[J]. Digital Manufacture Science, 2024, 2(22): 140-145 (in Chinese).
[15]	邓坚, 万小芳. 一种基于事件相机的缺陷检测装置: 中国, 217212305U[P]. 2022-08-16.
	DENG J, WAN X F. Defect detection device based on event camera: China, 217212305U[P]. 2022-08-16 (in Chinese).
[16]	YANG L X, ZHANG R Y, LI L D, et al. SimAM: a simple, parameter-free attention module for convolutional neural networks[EB/OL]. [2025-03-30]. https://proceedings.mlr.press/v139/yang21o.html.
[17]	GALLEGO G, DELBRUCK T, ORCHARD G, et al. Event-based vision: a survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(1): 154-180. DOI URL
[18]	CHEN S S, MARTINI B, CULURCIELLO E. A bio-inspired event-based size and position invariant human posture recognition algorithm[C]// 2009 IEEE International Symposium on Circuits and Systems. New York: IEEE Press, 2009: 775-778.
[19]	AYDIN A, GEHRIG M, GEHRIG D, et al. A hybrid ANN-SNN architecture for low-power and low-latency visual perception[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. New York: IEEE Press, 2024: 5701-5711.
[20]	FAN Q D. Research on the dynamic object segmentation based on computer vision technology[C]// 2024 IEEE 2nd International Conference on Sensors, Electronics and Computer Engineering. New York: IEEE Press, 2024: 924-927.
[21]	HE X, ZHAO D C, LI Y, et al. An efficient knowledge transfer strategy for spiking neural networks from static to event domain[C]// The 38th AAAI Conference on Artificial Intelligence. Washington: AAAI Press, 2023: 512-520.
[22]	SU Q Y, CHOU Y H, HU Y F, et al. Deep directly-trained spiking neural networks for object detection[C]// 2023 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2023: 6532-6542.
[23]	TIAN Y J, YE Q X, DOERMANN D. YOLOv12:attention- centric real-time object detectors[EB/OL]. [2025-02-18]. https://arxiv.org/abs/2502.12524.pdf.
[24]	KIM J, BAE J, PARK G, et al. N-ImageNet: towards robust, fine-grained object recognition with event cameras[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 2126-2136.
[25]	MA X, DAI X Y, BAI Y, et al. Rewrite the star[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 5694-5703.
[26]	GUO A J, MA J P, DIAN R W, et al. Surface defect detection competition with a bio-inspired vision sensor[J]. National Science Review, 2023, 10(22): 140-145.