【作者】 王浩；

【导师】 薛林；

【作者基本信息】 大连理工大学 ， 机械工程， 2025， 硕士

【摘要】 钢板作为工业领域最重要原材料和中间产品,因其超高的强度、良好的塑性和焊接性能,而在建筑、制造等多个领域发挥着不可替代的作用。然而钢板在加工过程中,容易受到加工工艺、原材料质量和加工设备的影响,产生一系列表面缺陷,这些缺陷将削弱钢板的强度、韧性和耐磨性,直接影响产品的质量和竞争力。因此,研发一种精度高、速度快的表面缺陷检测算法,具有极其重要的现实意义。本文对经典的Mask2former语义分割模型进行改进,提出了一种轻量化的钢板表面缺陷检测算法,通过针对性优化实现对钢板表面缺陷的快速定位与精确分割。本文主要研究内容如下:(1)针对钢板表面缺陷图像数据集样本不平衡的问题,提出一种复合式数据增强策略。使用传统数据扩增方式对数据进行增强,并在此基础上加入通过生成对抗网络生成的新图像。这种复合增强策略通过几何变换保持样本的空间一致性,同时利用生成对抗网络生成新的样本,不仅能有效平衡样本分布,还能提高缺陷形态多样性。(2)针对传统钢板表面缺陷检测算法流程复杂,需求算力高的问题,本文提出一种改进Mask2Former的轻量化缺陷检测算法。首先在骨干网络中引入Efficient Vit模块进行轻量化处理,采用级联分组注意力机制减少模型参数量,然后使用轻量的FFN层替换Masked-attention模块中的自注意力模块。实验结果表明,通过这两种优化方案,模型的参数量降低了37.98%。(3)针对传统钢板表面缺陷检测算法边缘分割精度不高的问题,本文使用空间特征交互模块融合不同深度特征图,进一步增强空间信息交换,并通过特征压缩结构模块将输出特征图统一;然后使用非对称统一焦点损失函数作为掩码损失函数替代原本的Combo Loss提升模型对复杂数据的适应能力。实验结果表明,通过这两种优化方案,模型的MIo U分别提高了2.54%和1.86%。(4)为了使普通工人快速掌握缺陷检测流程,本文开发了一套基于Docker的钢板表面缺陷检测系统,利用Web前端开发技术结合PHP设计了钢板缺陷检测系统的人机交互界面,并利用My SQL对检测结果进行云存档,从而实现语义分割模型的云端部署。更多还原

【Abstract】 Steel sheets,as one of the most critical raw materials and intermediate products in industrial applications,play an irreplaceable role in construction,manufacturing,transportation,and other sectors due to their exceptional strength,excellent plasticity,and weldability.However,during processing,steel plates are prone to surface defects induced by manufacturing techniques,raw material quality,and equipment conditions.These defects compromise their strength,toughness,and wear resistance,directly impairing product quality and market competitiveness.Therefore,developing a high-precision and rapid surface defect detection algorithm holds significant practical importance.Compared to traditional surface defect detection methods,semantic segmentation-based algorithms demonstrate notable advantages in segmentation accuracy and adaptability.This study improves the classical Mask2Former semantic segmentation model and proposes a lightweight algorithm for steel plate defect detection,achieving rapid localization and precise segmentation of surface defects through targeted optimization.The main contributions of this study are as follows:(1)To address the class imbalance issue in steel plate surface defect image datasets,a hybrid data augmentation strategy is proposed.Conventional data augmentation methods are first applied,followed by supplementing with synthetic images generated by generative adversarial networks(GANs),thereby enhancing the representation of rare samples in the dataset.This hybrid augmentation approach preserves spatial consistency through geometric transformations while generating novel samples via GANs,effectively balancing class distribution and increasing defect morphological diversity.(2)To address the issues of complex pipelines and high computational demands in conventional steel plate defect detection algorithms,this paper proposes a lightweight defect detection algorithm based on an improved Mask2Former architecture.The approach first incorporates Efficient Vi T modules into the backbone network for lightweight processing,employs cascaded group attention mechanisms to reduce model parameters,and subsequently replaces self-attention modules in the masked-attention blocks with lightweight FFN layers to streamline the architecture.Experimental results demonstrate that these dual optimization strategies achieve a 37.98%reduction in model parameters.(3)To address the insufficient edge segmentation accuracy in conventional steel plate defect detection algorithms,this study employs a Spatial Interaction Module(SIM)to fuse multi-depth feature maps,thereby enhancing spatial information exchange,while a Feature Compression Module(FCM)standardizes the output feature maps.An Asymmetric Unified Focal Loss is subsequently adopted as the mask loss function,replacing the original Combo Loss to improve the model’s adaptability to complex data patterns.Experimental results indicate that these dual optimizations yield MIo U improvements of 2.54%and 1.86%respectively.(4)To enable rapid adoption by non-expert operators,this study develops a Docker-based steel plate surface defect detection system,featuring a human-machine interface developed with web front-end technologies and PHP,with cloud-based result archiving via My SQL to enable cloud deployment of the semantic segmentation model.更多还原