ILN-YOLOv8: A Lightweight Image Recognition Model for Crimped Wire Connectors

• Published in: IEEE access 2025, Vol.13, p.5193-5202
• Main Authors: Zhou, Xiaojian, Kan, Jicheng, Fatin Liyana Mohd Rosely, Nur, Duan, Xu, Cai, Jiajing, Zhou, Zihan
• Format: Article
• Language: English
• Subjects: Accuracy; Attention mechanisms; Connectors; Convolution; Crimped wire connector; Crimping; Feature extraction; ILN-YOLOv8; Image recognition; knowledge distillation; lightweight; Object detection; Presses; pruning; Wire
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3525564

With the rapid growth of China's new energy vehicle industry, the quality of crimped wire connectors directly impacts the performance of wiring harnesses, which are critical to the overall vehicle quality. At present, reliable methods for inspecting crimped wire connectors are still primarily based on image recognition evaluations. To address this, we propose an Improved Lightweight Network based on YOLOv8 (ILN-YOLOv8) to achieve high-precision and high-efficiency detection of crimped wire connectors. Taking the original YOLOv8 model as a baseline, the new model enhances the ability to extract shallow features from small targets by increasing the P2 detection layer and improving the Feature Pyramid Network(FPN) and Path Aggregation Network(PAN) structures. Next, the improved Selective Boundary Aggregation(SBA) module replaces the Concat module in the Neck, enhancing the fusion of deep and shallow features. Additionally, the Efficient Local Attention(ELA) attention mechanism is introduced into the Cryptographic Service Provider(CSP) bottleneck with 2 convolutions(C2F) module in the Backbone, improving feature localization accuracy without increasing network complexity. The Minimum Point Distance based IoU(MPDIoU) loss function is used to calculate localization loss, improving detection accuracy while preventing gradient explosion. Finally, lightweighting of the ILN-YOLOv8 model is achieved using the slim-neck network, the backbone with Depthwise Separable Convolution (DWConv), and Lightweight Convolution (LightConv) modules. After pruning and knowledge distillation, the model's complexity and computational load significantly decreased while accuracy improved, meeting the industry's requirements for crimped wire connectors detection and achieves superior performance. Experimental results show that, compared to the original YOLOv8 model, the proposed method achieved 96.2% accuracy on a real-world crimped wire connectors dataset, with mAP@0.5 and mAP@.5:.95 improving by 6.1% and 7.4%, respectively, while Parameters and FLOPs decreased by 66.7% and 34.6%.