Enhanced YOLOv7 for EMU Damage Detection: Overcoming False Detection and Data Scarcity by Network Optimization and AIGC

Abstract

In the context of advancements in Artificial Intelligence Generated Content (AIGC) technology, this study focuses on addressing the challenges of data scarcity and high false negative rates in the detection of damage for Electric Multiple Units (EMU). To overcome these challenges, we propose an enhanced You Only Look Once version 7 (YOLOv7) model. First, we utilize the Low-Rank Adaptation (LoRA) lightweight training strategy, which utilizes a small number of actual damage images (such as foreign objects, oil leaks, and scratches). We also use the Stable Diffusion model to generate additional high-quality damage samples, effectively enriching the training dataset. Second, we incorporate the Coordinate Concatenated Spatial(CS) Attention mechanism into the YOLOv7 backbone network, adaptively adjusting channel and spatial attention weights to improve the model's ability to detect targets in complex backgrounds while maintaining a lightweight design. Third, we decouple the detection head to independently perform classification and localization tasks. Finally, we introduce the Focal-Efficient Intersection over Union (Focal-EIOU) loss function to optimize gradient allocation during training, promoting rapid convergence of high-quality anchor boxes and improving bounding box prediction accuracy. Experiments conducted on a dataset of 3717 generated images demonstrate that the improved YOLOv7 model achieves Mean Average Precision (mAP) and Recall rates of 98.0% and 96.1%, respectively, representing significant improvements over other YOLO models.