基于混合注意力机制的钢材表面缺陷分类算法

Classification of steel surface defects based on multi-attention mechanism

摘要:

缺陷检测在钢材生产中起着重要作用。针对钢材缺陷特点及当前分类算法对复杂模型与非启发式注意力机制的依赖,论文提出一种基于CBAM混合注意力架构的钢材表面缺陷分类算法(Multi-Attention based Classification Method, MACM)。首先,通过灰色关联度分析实现启发式通道注意力,增强算法可解释性同时克服层间信息丢失对通道注意力判定的影响。其次,采用紧凑双线性池化实现空间特征融合,增强抽象特征之间非线性复杂交互信息的捕获。最后基于CBAM架构实现通道和空间注意力的结合。实验表明,MACM算法在实现较好性能的同时相较现有SOTA算法更为轻量,验证了MACM算法在提升分类准确性、降低模型复杂度方面的有效性。

Abstract:

Defect detection plays an important role in steel production. In view of the characteristics of steel defects and the current classification methods’ dependence on complex models and non-heuristic attention mechanisms, this paper proposes a Multi-Attention based Classification Method (MACM) for steel surface defect classification based on the CBAM hybrid attention architecture. Firstly, heuristic channel attention is achieved through gray correlation analysis to enhance the interpretability of the algorithm and overcome the impact of inter-layer information loss on channel attention determination. Secondly, compact bilinear pooling is used to achieve spatial feature fusion, enhancing the capture of nonlinear complex interaction information between abstract features. Finally, the combination of channel and spatial attention is achieved based on the CBAM architecture. Experiments show that the MACM method achieves good performance while being lighter than existing SOTA methods, validating the effectiveness of the MACM algorithm in improving classification accuracy and reducing model complexity.