Treffer: 基于轻量化卷积块注意力 Transformer 的 表面缺陷检测算法.
Weitere Informationen
To address the issues of small differences between defects and background, as well as large variations within the same class of defects in defect semantic segmentation tasks, this paper proposes a surface defect detection algorithm based on Lightweight Convolutional Block Attention Transformer, named LCBAFormer. The proposed algorithm aims to enhance the accuracy of segmentation for various types of defects. Firstly, a Lightweight Convolutional Block Attention Module (LCBAM) is designed, which integrates channel attention and spatial attention modules to extract effective channel and spatial information. This enables the model to focus more on local defect feature information, enhance feature differences between defects, and mitigate variations within the same category of defects. Secondly, a lightweight Semantic Injection Module (SIM) is introduced, which gradually fuses multi-scale feature information and thereby improving the network's ability to locate and distinguish different defects. The experimental results show that on the NEU-Seg steel strip defect dataset and the magnetic tile defect dataset (MT-Defect), the proposed algorithm achieves mean Intersection over Union (mIoU) of 84. 75% and 79. 46%, mean Recall (mRec) rates of 92. 29% and 87. 50%, and mean F1 scores (mF1) of 91. 52% and 88. 08%, respectively. Additionally, the algorithm exhibits low computational complexity, with 1. 03 and 2. 65 GFLOPs (billion floating-point operations per second) on the NEU-Seg and MT-Defect datasets, respectively. Compared to mainstream methods, the proposed algorithm features fewer parameters and superior segmentation results, achieving a good balance between parameter count and detection performance. [ABSTRACT FROM AUTHOR]
针对缺陷语义分割任务中缺陷与背 景差异小以及同类缺陷间差异大的问 题, 本文提出一种基于轻量化卷积块注 意力 Transformer 的表面缺陷检测算法 (LCBAFormer),旨在提高不同类型缺陷 的分割准确率. 首先, 设计轻量化卷积块 注意力模块 LCBAM, 通过结合通道注意 力模块和空间注意力模块, 提取有效的 通道信息和空间信息, 使模型更专注于 局部缺陷特征信息, 增强不同缺陷间的 特征差异, 减少同类缺陷间的差异. 其 次, 提出轻量化的语义注入模块 SIM, 以 逐步融合多尺度特征信息, 提升网络对 不同缺陷的定位和区分能力. 在 NEUSeg 钢带缺陷数据集和 MT-Defect 磁瓦缺 陷数据集上的实验结果表明: 本文提出 的 算 法 平 均 交 并 比 (mIoU) 分 别 为 84. 75%和 79. 46%; 平均召回率 (mRec) 分别为 92. 29%和 87. 50%; 平均 F 1 分数 (mF1) 分别为 91. 52% 和 88. 08%; 算法 的计算复杂度较低, 每秒 10 亿次浮点运 算次数 (GFLOPs) 分别为 1. 03 和 2. 65. 与主流算法相比, 本文方法在参数量和 检测性能上做到了较好的平衡, 具有更 小的参数量和更好的分割结果. [ABSTRACT FROM AUTHOR]