材料科学
曲面(拓扑)
复合材料
焊接
噪音(视频)
变形(气象学)
冶金
有限元法
过程(计算)
结构工程
出处
期刊:
日期:2026-05-15
卷期号:: 847-850
标识
DOI:10.1109/ainit70033.2026.11558016
摘要
Steel surface defect detection is limited by insufficient representation of minute defect features, difficulty in capturing irregular shapes, and low inference efficiency in complex backgrounds. To address these, this paper proposes an improved RT-DETR algorithm. It integrates a UniConvBlock-based cross-stage local receptive field aggregator into the backbone network, using large-kernel convolutions to build Gaussian-like receptive fields and enhance global context capture. In the AIFI neck network, a single-head self-attention mechanism with parallel perceptual branches reduces computational redundancy and memory access overhead while maintaining expressive power. Additionally, a matching perceptual loss function dynamically adjusts penalty weights, strengthening supervision of low-quality samples and alleviating convergence issues under sparse queries. Results show the model achieves 74.04% mAP50 and 74.60% precision on the NEU-DET dataset, improving by 1.24% and 1.60% over the baseline with 15.17M parameters. This algorithm reduces complexity while improving accuracy, meeting industrial edge device requirements for real-time, high-precision detection.
科研通智能强力驱动
Strongly Powered by AbleSci AI