基于轻量化YOLOv4的黏稠食品灌装成品缺陷检测

针对黏稠流质食品的自动化灌装中,传统灌装成品缺陷检测方法难以同时对多目标高速检测的难题,提出一种基于YOLOv4目标检测算法的轻量级灌装成品缺陷检测方法.MobileNetV3主干特征提取网络能对输入样本进行轻量级特征提取;增强特征提取网络采用深度可分离卷积策略,以降低参数计算量,然后通过设计的全面路径聚合网络(FPANet)和引入的通道注意力机制(ECA)提升增强特征提取网络对于目标特征的靶向表达.将设计的轻量化网络进行模型训练和精度测试,并在同一数据集下与其它目标检测算法进行对比,以分析本文方法的优劣.实验结果表明,本文方法能够在保持精度的前提下提升检测速度,实现了黏稠食品灌装成品缺陷的多目标高速检测.

Defect Detection in Filling Products of Viscous Food Based on Lightweight YOLOv4

The lightweight filling product defect detection method based on YOLOv4 object detection algorithm was proposed aimed at the issue of difficulty in detecting multiple targets at the same time in the automatic filling of viscous liquid food with the traditional filling product defect detection method. The lightweight feature extraction was performed on the input samples through the MobileNetV3 backbone feature extraction network, and the depthwise separable convolution strategy was used to reduce the computational cost of the enhance feature extraction network. Then, the full path aggregation network (FPANet) was designed and efficient channel attention (ECA) mechanism was introduced to improve the target feature expression of the enhance feature extraction network. The model training and precision testing were carried out on the designed lightweight network, and the performance of other object detection algorithms was compared in the same dataset to reveal the superiority and disadvantages of the proposed method. The experimental results showed that the proposed method could improve the detection speed while maintaining the accuracy, and the multi-objective high-speed detection of the filling product defects of viscous food was realized.