LIDD-Net:基于深度学习的轻量级工业产品缺陷检测方法

doi:10.12068/j.issn.1005-3026.2025.20240058

摘要/Abstract

摘要：

工业产品中存在各种缺陷，且不同类型缺陷之间存在着高度相似、尺度变化大、背景信息复杂等问题.为解决这些问题，本文提出了轻量级工业缺陷检测网络（LIDD-Net）.针对相似度高的不同种类缺陷，LIDD-Net设计了通道交互分离骨干网络，在降低模型计算量的同时提高了特征提取能力；针对不同尺度的缺陷，LIDD-Net提出了轻量化特征融合网络RepGhostPAN，在能融合图像中多尺度特征的同时提高了推理速度；针对检测背景的复杂性，LIDD-Net提出了轻量辅助训练模块，通过使用辅助训练头和动态软标签分配策略，可更好地区分目标缺陷与复杂背景.通过在钢材缺陷、铝缺陷和轮胎缺陷数据集的实验结果表明，LIDD-Net在参数量仅为0.62×10⁶的情况下分别获得了98.3%，98.1%和96.1%的mAP@0.5，可以满足工业现场实际需求.

关键词: 工业缺陷检测, 轻量化检测网络, 特征融合, 结构重参数化, 注意力机制

Abstract:

In industrial products， various types of defects often exhibit high inter-class similarity， large scale variations， and complex backgrounds. To address these challenges， a lightweight industrial defect detection network （LIDD-Net） was proposed. To handle highly similar defect types， in LIDD-Net， a channel interaction separation backbone network was introduced， which enhanced feature extraction while reducing the computational cost of the model. To address multi-scale defect variations， a lightweight feature fusion network was developed， namely RepGhostPAN， to efficiently integrate multi-scale features in the image and accelerate inference. For complex detection backgrounds， a lightweight auxiliary training module was proposed， leveraging an auxiliary training head and a dynamic soft label assignment strategy to better distinguish target defects from complex backgrounds. Experiments on steel， aluminum， and tire defect datasets demonstrate that LIDD-Net achieves mAP@0.5 scores of 98.3%， 98.1%， and 96.1%， respectively， with only 0.62×10⁶ parameters， meeting practical industrial requirements.

Key words: industrial defect detection, lightweight detection network, feature fusion, structural reparameterization, attention mechanism

中图分类号:

TP 391.4

沙晓鹏, 谢德瀚, 郭周鹏, 孙凯. LIDD-Net:基于深度学习的轻量级工业产品缺陷检测方法[J]. 东北大学学报（自然科学版）, 2025, 46(10): 18-26.

Xiao-peng SHA, De-han XIE, Zhou-peng GUO, Kai SUN. LIDD-Net: Lightweight Industrial Product Defect Detection Method Based on Deep Learning[J]. Journal of Northeastern University(Natural Science), 2025, 46(10): 18-26.

图/表 14

参考文献 12

[1]	Shen X K， Liu J H， Jiang L， et al. A novel weld defect detection method for intelligent magnetic flux leakage detection system via contextual relation network［J］. IEEE Transactions on Industrial Electronics， 2024， 71（6）： 6304-6314.
[2]	Zong Z F， Song G L， Liu Y. DETRs with collaborative hybrid assignments training［C］// Proceedings of the IEEE Conference on International Conference on Computer Vision. Paris， 2023： 6725-6735.
[3]	Yu Z P， Huang H B， Chen W J， et al. YOLO-FaceV2： a scale and occlusion aware face detector［J］. Pattern Recognition， 2024， 155： 110714.
[4]	Ding X H， Zhang X Y， Ma N N， et al. RepVGG： making VGG-style convnets great again［C］ // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Virtual， 2021： 13733-13742.
[5]	Lei Y， Chen Y R， Hai T， et al. DGNet： an adaptive lightweight defect detection model for new energy vehicle battery current collector［J］. IEEE Sensors Journal， 2023， 23： 29812-29830.
[6]	Ye L Q. AugShuffleNet： communicate more， compute less ［EB/OL］. （2022-03-13）［2022-08-21］. .
[7]	Yang L X， Zhang R Y， Li L D， et al. SimAM： a simple， parameter-free attention module for convolutional neural networks［C］// International Conference on Machine Learning. Virtual， 2021： 11863-11874.
[8]	Cao J M， Li Y Y， Sun M C， et al. Do-Conv： depthwise over-parameterized convolutional layer［J］. IEEE Transactions on Image Processing， 2022， 31： 3726-3736.
[9]	Filippas D， Nicopoulos C， Dimitrakopoulos G. Streaming dilated convolution engine［J］. IEEE Transactions on Very Large Scale Integration （VLSI） Systems， 2023， 31： 401-405.
[10]	Shamsolmoali P， Chanussot J， Zareapoor M， et al. Multipatch feature pyramid network for weakly supervised object detection in optical remote sensing images［J］. IEEE Transactions on Geoscience and Remote Sensing， 2022， 60： 1-13.
[11]	Han K， Wang Y H， Tian Q， et al. Ghostnet： more features from cheap operations［C］// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Virtual， 2020： 1580-1589.
[12]	Mishra R， Gupta H P. Designing and training of lightweight neural networks on edge devices using early halting in knowledge distillation［J］. IEEE Transactions on Mobile Computing， 2024， 23： 4665-4677.

操作系统	Microsoft Windows 11
CPU	Intel Core i5-13600
GPU	NVIDIA GeForce RTX 4090，CUDA=12.2
运行环境	Python 3.8，PyTorch 1.13

操作系统	Microsoft Windows 11
CPU	Intel Core i5-13600
GPU	NVIDIA GeForce RTX 4090，CUDA=12.2
运行环境	Python 3.8，PyTorch 1.13

模型	mAP@0.5/%	mAP@0.5:0.95/%	Params×10^-6	FLOPs×10^-9
BaseLine	95.37	64.78	1.171 8	0.163 2
+LATM	96.68	66.76	1.171 8	0.163 2
+CISB-Net	98.06	72.42	0.620 3	0.109 7
+RepGhostPAN	98.30	73.26	0.628 7	0.110 4

模型	mAP@0.5/%	mAP@0.5:0.95/%	Params×10^-6	FLOPs×10^-9
BaseLine	95.37	64.78	1.171 8	0.163 2
+LATM	96.68	66.76	1.171 8	0.163 2
+CISB-Net	98.06	72.42	0.620 3	0.109 7
+RepGhostPAN	98.30	73.26	0.628 7	0.110 4

网络	mAP@0.5/%	mAP@0.5∶0.95/%	Params×10^-6	FLOPs×10^-9
YOLOv5n	93.28	59.10	1.772 0	0.382 1
YOLOv8n	90.43	57.93	3.006 8	0.738 0
YOLO-Fastestv2	88.02	55.72	0.237 6	0.003 4
YOLOX-Nano	95.92	61.14	0.897 7	0.114 9
PP-PicoDet-S	93.44	64.12	0.989 4	0.462 8
Nanodet	94.02	61.64	0.928 5	0.124 5
Nanodet-plus	96.62	66.79	1.172 6	0.166 5
TOOD	89.37	63.42	31.234	6.359 3
Foveabox	90.76	56.53	36.244	7.129 8
LIDD-Net	98.30	73.26	0.628 7	0.110 4