关键词: 织物疵点检测, 深度学习, 卷积神经网络, 傅里叶变换, 频域卷积, 织物图像

Abstract:

Objective This study aims to developing an advanced frequency-domain convolution module to overcome the limitations in texture recognition and feature representation in textile defect detection. By effectively integrating two-dimensional Fourier transform with conventional convolution, this research harnesses the periodic characteristics of fabric images, decouples global and local features, and enhances the model's ability to represent textile image features. The importance and necessity of this study arise from the need for more efficient and accurate textile quality control.
Method Combining theoretical models with empirical methods, the proposed approach integrates the advantages of two-dimensional Fourier transform with deep learning. A specialized frequency-domain convolution module was designed specifically for textile image defect detection and applied within the YOLOv5 object detection framework. This study explores the application of image frequency-domain representation in textiles, using frequency-domain convolution to enhance the global receptive field, improve model performance, and ensure computational efficiency. By organically combining large and small kernel convolution techniques, the model architecture is optimized to decouple local and global feature processing, enhancing the analysis process of textile image features. By incorporating prior experience, the representation of frequency-domain features is optimized, improving the model's interpretability, reducing the learning burden, and enhancing feature expression capabilities. Additionally, an attention mechanism component designed for frequency-domain features is implemented to optimize the extraction of these features.
Results It demonstrates the performance enhancements achieved by different models in textile defect detection across various datasets. The models integrating frequency-domain convolution have shown significant improvements, especially in more complex datasets. For instance, the proposed model 1 displayed performance increments in m_AP50 (mean average precision at an IoU threshold 50%) by 1.5%, 5.0%, 11.9%, and 16.3% and in m_AP75 (mean average precision at a stricter IoU threshold of 75%) by 6.8%, 19.1%, 28.9%, and 32.9% across the datasets, indicating the effectiveness of incorporating frequency-domain convolution to address the challenges in textile image processing. This method leverages the global receptive field and decouples global and local features, thereby enhancing the model's capability to process diverse fabric types and defect patterns effectively. Additionally, the proposed model 1 showed superior detection speed (189.1 frames per second) compared to the baseline YOLOv5 model (158.7 frames per second), confirming that the computational efficiency of frequency-domain convolution excels in shallow network layers because of faster matrix multiplication speeds, despite a decrease in efficiency in deeper layers where Fourier transformations become costly. This structure proves particularly advantageous for textile defect detection, where the complexity and variety of fabric types and defects require robust and adaptable models.Extensive ablation experiments were conducted to validate the efficacy of the frequency-domain convolution module (FFC-tex), specifically optimized for textile image processing. In ablation model 1, a pure convolution module with an FFC structure was adopted to replace the FFC-tex module to isolate the impact of the feature crossover structure. Ablation model 2 integrated the complete original FFC module to confirm the effectiveness of frequency-domain convolution. In ablation model 3, frequency-domain features were restructured, including the decoupling of amplitude and phase features and the creation of frequency distribution and waveform orientation maps, so as to validate the effectiveness of these components. Ultimately, by comparing the proposed model 1 with the ablation model 3, the effectiveness of the introduced frequency-domain attention mechanism was examined. The significant improvements in m_AP50 scores during the addition of these components underscore the necessity of each component in the FFC-tex module for enhancing model performance. The optimized FFC-tex module effectively utilizes the frequency-domain features of fabric images, showing notable improvements in precision and generalization capabilities compared to baseline and earlier ablation models, making it suitable for a variety of fabric types and defect patterns.
Conclusion The YOLOv5 model is improved by integrating frequency-domain convolution, improving its generalization and robustness for textile defect detection. Frequency-domain convolution provides a global receptive field, allowing shallow network layers to utilize more contextual information and better understand complex structures. It simplifies computational complexity by transforming time-domain convolutions into element-wise multiplications in the frequency domain. The optimized FFC-tex module enhances feature extraction and generalization across fabric types, significantly boosting model performance in defect detection by decoupling local and global features.

Key words: fabric defect detection, deep learning, convolutional neural network, Fourier transform, frequency-domain convolution, fabric image