关键词: 梳棉机棉网, 深度学习, 目标检测, 棉结, 轻量化模型, YOLOv8, 图像检测

Abstract:

Objective In order to improve the nep detection using deep learning with minimal computing resources and to achieve real-time online detection using embedded devices and mobile terminals, this research proposed an improved YOLOv8 detection method.

Method Lightweight network MobileNetv3_Small was used as YOLOv8 backbone network so as to reduce the number of parameters. The self-improved CA (coordinate attention) model was adopted to replace the SE (squeeze-and-excitation)attention mechanism model in MobileNetv3 so as to improve the accuracy of nep detection. EIoU loss function substituted CIoU loss function was employed to retain more effective information in data processing.

Results The dataset constructed includes three target objects, which were named lumpy nep, strip nep and cotton miscellaneous matters. Different light intensities and camera orientation on model detection effec was studied by using test set. The experimental results show that no missed detections and false detections of all target objects was found under different light intensities and different camera orientations. The average accuracy of improved YOLOv8 model reached 95.8%, with an increase of 2.6% compared with the reported model. Compared with the improved YOLOv8, YOLOv7, YOLOv5 and Faster R-CNN network models, the improved YOLOv8 network model involved fewer parameters but with higher average accuracy. According to the average accuracy of the improved YOLOv8 model for each type, the detection rate reached more than 95.1% with high accuracy. The improved model ablation experiments show when MobileNetv3 replaced the YOLOv8 backbone network, the parameter quantity is greatly reduced, and it was found from the comparison experiment between the loss function CIoU and EIoU that when the loss function is EIoU, the average accuracy of model detection is higher than CIoU. In addition, the number of parameters in the model became smaller, which is beneficial for embedded equipment and would satisfy the actual industrial production requirements.

Conclusion A nep detection method based on improved YOLOv8 model is proposed combining the characteristics of nep. By improving the original YOLOv8 model and inducting improved MobileNetv3 module, the accuracy of the model is improved, and the number of parameters is reduced, which benefits effective detection. The improved model can effectively reduce the parameter number and make the detection accuracy improved. Compared with the classic detection methods, the accuracy of the proposed model is higher, and the parameters get fewer. This method can be used with mobile devices or built-in devices for intelligent factory equipment.

Key words: carding net, deep learning, object detection, nep, lightweight model, YOLOv8, image detection