Li Min, Ye Pei, Cui Shuqin, Zhu Ping, Liu Junping
School of Computer and Artificial Intelligence, Wuhan Textile Unversity, Wuhan 430200, China.
Sensors (Basel). 2024 Dec 21;24(24):8181. doi: 10.3390/s24248181.
Currently, fabric defect detection methods predominantly rely on CNN models. However, due to the inherent limitations of CNNs, such models struggle to capture long-distance dependencies in images and fail to accurately detect complex defect features. While Transformers excel at modeling long-range dependencies, their quadratic computational complexity poses significant challenges. To address these issues, we propose combining CNNs with Transformers and introduce Kolmogorov-Arnold Networks (KANs) to enhance feature extraction capabilities. Specifically, we designed a novel network for fabric defect segmentation, named HKAN, consisting of three components: encoder, bottleneck, and decoder. First, we developed a simple yet effective KANConv Block using KAN convolutions. Next, we replaced the MLP in PoolFormer with KAN, creating a lightweight KANTransformer Block. Finally, we unified the KANConv Block and the KANTransformer Block into a Hybrid KAN Block, which serves as both the encoder and bottleneck of HKAN. Extensive experiments on three fabric datasets demonstrate that HKAN outperforms mainstream semantic segmentation models, achieving superior segmentation performance and delivering prominent results across diverse fabric images.
目前，织物缺陷检测方法主要依赖于卷积神经网络（CNN）模型。然而，由于卷积神经网络的固有局限性，这类模型难以捕捉图像中的长距离依赖关系，并且无法准确检测复杂的缺陷特征。虽然Transformer在对长距离依赖关系进行建模方面表现出色，但其二次计算复杂度带来了重大挑战。为了解决这些问题，我们提出将卷积神经网络与Transformer相结合，并引入柯尔莫哥洛夫 - 阿诺德网络（KANs）以增强特征提取能力。具体而言，我们设计了一种用于织物缺陷分割的新型网络，名为HKAN，它由三个部分组成：编码器、瓶颈层和解码器。首先，我们使用KAN卷积开发了一个简单而有效的KANConv模块。接下来，我们用KAN替换了PoolFormer中的多层感知器（MLP），创建了一个轻量级的KANTransformer模块。最后，我们将KANConv模块和KANTransformer模块统一为一个混合KAN模块，它同时作为HKAN的编码器和瓶颈层。在三个织物数据集上进行的大量实验表明，HKAN优于主流语义分割模型，在各种织物图像上实现了卓越的分割性能并取得了显著成果。
