Zhou Wujie, Sun Xinyu, Qian Xiaohong, Fang Meixin
IEEE Trans Neural Netw Learn Syst. 2025 Jul;36(7):12469-12482. doi: 10.1109/TNNLS.2024.3479453.
Owing to extensive research on deep learning, significant progress has recently been made in trackless surface defect detection (SDD). Nevertheless, existing algorithms face two main challenges. First, while depth features contain rich spatial structure features, most models only accept red-green-blue (RGB) features as input, which severely constrains performance. Thus, this study proposes a dual-stream teacher model termed the asymmetrical contrastive learning network (ACLNet-T), which extracts both RGB and depth features to achieve high performance. Second, the introduction of the dual-stream model facilitates an exponential increase in the number of parameters. As a solution, we designed a single-stream student model (ACLNet-S) that extracted RGB features. We leveraged a contrastive distillation loss via knowledge distillation (KD) techniques to transfer rich multimodal features from the ACLNet-T to the ACLNet-S pixel by pixel and channel by channel. Furthermore, to compensate for the lack of contrastive distillation loss that focuses exclusively on local features, we employed multiscale graph mapping to establish long-range dependencies and transfer global features to the ACLNet-S through multiscale graph mapping distillation loss. Finally, an attentional distillation loss based on the adaptive attention decoder (AAD) was designed to further improve the performance of the ACLNet-S. Consequently, we obtained the ACLNet-S*, which achieved performance similar to that of ACLNet-T, despite having a nearly eightfold parameter count gap. Through comprehensive experimentation using the industrial RGB-D dataset NEU RSDDS-AUG, the ACLNet-S* (ACLNet-S with KD) was confirmed to outperform 16 state-of-the-art methods. Moreover, to showcase the generalization capacity of ACLNet-S*, the proposed network was evaluated on three additional public datasets, and ACLNet-S* achieved comparable results. The code is available at https://github.com/Yuride0404127/ACLNet-KD.
由于对深度学习的广泛研究，最近在无轨表面缺陷检测（SDD）方面取得了重大进展。然而，现有算法面临两个主要挑战。首先，虽然深度特征包含丰富的空间结构特征，但大多数模型只接受红绿蓝（RGB）特征作为输入，这严重限制了性能。因此，本研究提出了一种双流教师模型，称为不对称对比学习网络（ACLNet-T），它同时提取RGB和深度特征以实现高性能。其次，双流模型的引入使得参数数量呈指数级增长。作为一种解决方案，我们设计了一种提取RGB特征的单流学生模型（ACLNet-S）。我们通过知识蒸馏（KD）技术利用对比蒸馏损失，将丰富的多模态特征逐像素、逐通道地从ACLNet-T转移到ACLNet-S。此外，为了弥补仅关注局部特征的对比蒸馏损失的不足，我们采用多尺度图映射来建立长程依赖关系，并通过多尺度图映射蒸馏损失将全局特征转移到ACLNet-S。最后，设计了一种基于自适应注意力解码器（AAD）的注意力蒸馏损失，以进一步提高ACLNet-S的性能。因此，我们得到了ACLNet-S*，尽管其参数数量差距近八倍，但性能与ACLNet-T相似。通过使用工业RGB-D数据集NEU RSDDS-AUG进行的综合实验，证实ACLNet-S*（带有KD的ACLNet-S）优于16种最新方法。此外，为了展示ACLNet-S的泛化能力，在所提出的网络在另外三个公共数据集上进行了评估，并且ACLNet-S取得了可比的结果。代码可在https://github.com/Yuride0404127/ACLNet-KD获取。
