LwDetFormer: A Hybrid Mobile and Transformer-Based Deep Learning Model for Lockwire Object Detection

Abstract

Lockwire plays a crucial role in various industries, including aviation, automotive, power, oil, and gas, by ensuring the secure fastening of critical components. However, detecting Lockwire objects poses significant challenges due to their reflective metallic materials and surface textures that closely resemble the surrounding environment. Traditional object detection methods struggle with these characteristics, especially when working with small datasets and variable lighting conditions. To address these issues, we propose LwDetFormer, a hybrid CNN-Transformer model designed for high-precision detection of Lockwire objects in complex backgrounds. LwDetFormer integrates MobileNet's local feature extraction capabilities with Transformer's global feature modeling abilities, enhancing accuracy and robustness. The model includes innovative modules such as Spatial Pyramid Pooling Focus, Feature Enhancement Module, Feature Fusion Module, and Spatial Context Awareness Module. Experimental results on a test set containing 2,000 images show that LwDetFormer outperforms advanced models like YOLOv8, MobileNetv3, and MobileFormer in terms of precision and recall, achieving a precision rate of 95.9% and a recall rate of 92.0%. These findings highlight LwDetFormer's potential for improving safety and efficiency in the inspection of aircraft engine fuse parts.