关键词: 蕾丝织物, 深度学习, 表面缺陷检测, 轻量级模型, 注意力机制

Abstract:

Objective To ensure the quality of lace products, it is of great significance to achieve accurate and efficient surface defect detection during the lace fabric production process. To further reduce the computational complexity of the lace fabric surface defect detection models based on deep learning and make them more suitable for embedded devices with low computing power, a lightweight detection method named MosYOLO is proposed by improving the YOLOv9s model.

Method The improved MobileNetv3-Small lightweight network was adopted as the backbone network of YOLOv9s model to reduce the number of parameters and the amount of calculation. The improved Efficient Channel Attention mechanism was introduced into the backbone network to enhance the model's ability to recognize defect features. Aiming at the imbalanced proportion between the difficult and easy samples in the training data, the Focaler-CIoU loss was introduced to replace the original CIoU loss. The Focal Modulation module was applied to the neck network to further enhance the model's ability to extract defect information.

Results In this study, a dataset was constructed using lace fabric images collected from real industrial sites, including four types of surface defects, namely jacquard holes, broken yarns, holes and edging. The image size was all 512 pixels × 512 pixels. This dataset was used to train and test the model. By introducing multiple evaluation indicators for a comprehensive assessment of the model performance, it was learnt from the experiments that the mean Average Precision of improved MosYOLO model reached 91.0%, and the F1 score reached 88.4%. Compared with the baseline YOLOv9s model, mean Average Prescision and F1 score increased by 1.5% and 1.3%, respectively. Moreover, both the number of parameters and the amount of calculation of the model decreased by 27.1%, and the detection speed reached 37.9 frames per second. From the visualization results, the detection effect of MosYOLO model was superior to that of YOLOv9s for lace fabric surface defect detection, and it showed stronger detection ability for small defect. Compared with the Faster R-CNN, SSD, YOLOv4-Tiny, MobileNetv2-SSDLite, YOLOv7-Tiny and YOLOv7 object detection models, MosYOLO model achieved a better balance among detection accuracy, parameter quantity, computational cost and detection speed. By optimizing and accelerating MosYOLO model using TensorRT framework, the detection speed of the model and its deployment performance on edge device have been greatly improved. The ablation experiment results of improved model showed that after replacing the backbone network of YOLOv9s with the improved MobileNetv3-Small, the number of parameters, computational cost and volume of the model were significantly reduced. After introducing the Efficient Channel Attention mechanism in the shallow network stage, the mean Average Precision was improved. When the Focaler-CIoU bounding box regression loss was adopted, the mean Average Precision of the model became better than that of the CIoU loss. After using the lightweight Focal Modulation module in the neck network, the computational complexity of the model was further reduced, and the mean Average Precision was improved, making the model more efficient and more suitable for deployment on edge device.

Conclusion A lightweight method MosYOLO based on improved YOLOv9s model is proposed for the detection of lace fabric surface defects. The MosYOLO method significantly reduces the number of parameters, computational cost and volume of the model while ensuring the detection accuracy. By replacing the backbone network of YOLOv9s model with the improved MobileNetv3-Small, the improved Efficient Channel Attention mechanism, Focaler-CIoU bounding box regression loss and Focal Modulation module are introduced. MosYOLO outperforms YOLOv9s and other mainstream object detection models in multiple indicators such as the mean Average Precision and detection speed. The MosYOLO method can be deployed in edge device, better meeting the demands of the lace fabric industrial production site and enhancing production efficiency and product quality.

Key words: lace fabric, deep learning, surface defect detection, lightweight model, attention mechanism