一个基于SCCI-YOLO的高效轻量级算法,用于检测钢的表面缺陷
Huixiang Zhou1, Hong Zou2, Gaojun Hu1
1Department of Software Engineering, School of Software, East China Jiaotong University, No. 808 Shuanggang East Street, Nanchang, 330013, Jiangxi, China.
本研究介绍了SCCI-YOLO,这是一个改进的深度学习模型,用于检测钢表面缺陷. 它增强了特征提取和融合,大大提高了准确性,并减少了识别各种工业材料缺陷的错误.
科学领域:
- 材料科学 材料科学 材料科学
- 计算机视觉 计算机视觉
- 人工智能的人工智能
背景情况:
- 钢材表面缺陷对工业产品质量产生重大影响.
- 现有的深度学习方法在多样化,复杂化和多层次缺陷检测方面扎,导致不准确.
- 局限性包括缺陷的特征提取,融合和多尺度识别.
研究的目的:
- 开发一个先进的钢表面缺陷检测算法,克服当前的深度学习限制.
- 为了提高准确性,降低错误检测率,并增强识别多层次缺陷.
- 提出一个新的模型,SCCI-YOLO,基于改进的YOLOv8n架构.
主要方法:
- 将SPD-Conv模块集成到骨干中,用于自适应的卷积内核聚焦,增强小物体检测.
- 开发了C2f_EMA模块,用于改进特征提取和融合.
- 在Neck网络中集成了一种轻量级的跨尺度特征融合模块 (CCFM),以实现多尺度的适应性.
- 利用Inner-IoU损失函数来提高模型的收性和回归精度.
主要成果:
- 在NEU-DET数据集上,SCCI-YOLO实现了平均平均精度 (mAP) 的78.6%.
- 与YOLOv8n和YOLOv7相比,检测准确性分别提高了2.2%和5.9%.
- 与原始YOLOv8n模型相比,模型参数减少了43.9%.
结论:
- 在钢表面缺陷检测方面,SCCI-YOLO算法显示出卓越的整体性能.
- 提议的改进有效地解决了特征提取,融合和多尺度识别方面的挑战.
- 该模型为工业钢材表面质量控制提供了更准确,更有效的解决方案.
更多相关视频
11:34Subsurface Defect Localization by Structured Heating Using Laser Projected Photothermal Thermography
Published on: May 15, 2017
11:47Characterization of Surface Modifications by White Light Interferometry: Applications in Ion Sputtering, Laser Ablation, and Tribology Experiments
Published on: February 27, 2013
相关概念视频
Mechanical Characteristics of Steel
The tension test is fundamental for determining tensile strength. In this test, a steel specimen is stretched using a gripping device until it breaks. The data collected during this test are used...
Non-destructive Tests for Concrete Strength
Steel Manufacturing
During this smelting process, limestone plays a crucial role by forming slag. Slag captures impurities within the molten iron, such...
Response Surface Methodology
The process of RSM involves several key steps:
Yield Criteria for Ductile Materials under Plane Stress
The Maximum Shearing Stress Criterion, also known as...
Structural Steel Products
Once shaped, the steel's final form emerges as a continuous length, which is then segmented by a hot saw into manageable pieces. These segments...
