Journal of Electronic Imaging, Journal Year: 2024, Volume and Issue: 33(04)

Published: July 16, 2024

Prefabricated steel pipes play a crucial role in prefabricated buildings, and maintaining their surface integrity is to ensuring the safety of these buildings. We propose defect detection algorithm for pipes, D-YOLOv7-tiny, based on YOLOv7-tiny, address challenges high parameter count large computational requirements associated with traditional algorithms, making it difficult deploy at resource-constrained terminals. By incorporating squeeze-and-excitation attention mechanism into backbone network, D-YOLOv7-tiny effectively minimizes impact redundant information improves network's ability extract features. In addition, distribution shifting convolution implemented as replacement portion original effective layer aggregation network module network. This exchange reduces workload model without affecting its expressive power. Subsequently, lightweight ubiquitous content-aware reassembly features upsampling operator improved feature merging Finally, an attention-based dynamic head was adopted enhance model's robustness while minimizing counts. Compared mAP enhanced by 1.9% through experiments conducted self collected datasets. During this process, number parameters complexity decreased 7.6% 39.4%, respectively. The results show that method achieves meets practical engineering accuracy real-time performance.

Language: Английский