CLFI-YOLOv8s: An accurate and efficient model for bellows crack detection in air spring

Abstract

Cracks in air spring bellows significantly impact their service life. However, the surface cracks of bellows often exhibit low contrast, poor image quality, and complex backgrounds. Traditional detection methods struggle to achieve high precision and efficient crack identification. To address this issue, this paper proposes a CLFI-YOLOv8s model specifically designed for detecting cracks in bellows. Firstly, the convolutional priority multi-space (CPMS) attention module is integrated into the backbone to refine multi-scale feature extraction and localization. Subsequently, the C2f-LarK module in the neck expands the receptive field with large kernels, thereby improving spatial perception and fine-grained feature capture. To optimize efficiency, Partial Convolution (PConv) is integrated into the head, forming the Faster Detect structure, which reduces computational cost while maintaining detection accuracy. Additionally, Inner-Shape IoU replaces CIoU to further improve detection accuracy and generalization. Experimental results demonstrate that the CLFI-YOLOv8s outperforms the YOLOv8s in detection performance. The model achieves improvements in precision, recall, [email protected], and [email protected]–0.95 by 3.3 %, 3.5 %, 1.8 %, and 6.1 %, respectively. Simultaneously, the weight size, parameters, and GFLOPs are reduced by 14.4 %, 14.2 %, and 26.4 %, respectively. The results indicate that the CLFI-YOLOv8s model excels in crack detection tasks, demonstrating significant potential and practical value as a monitoring tool for air spring bellows cracks.