小尺度特征提取的瓶装矿泉水异物检测方法

瓶装矿泉水异物检测旨在查找出矿泉水中的异物杂质,在矿泉水制造业具有重要的应用价值。然而,传统的机器视觉检测精度较低,受背景的干扰较大、且漏检率和误检率较高。为解决上述问题,提出小尺度特征提取的瓶装矿泉水异物检测算法。该算法主要包含判别性特征学习模块,数据增值模块以及细粒度信息获取模块。在判别性特征学习模块中,针对矿泉水异物,利用聚类的方式设计合理的先验框尺寸并对特征图进行处理,通过对网络的输出结果进行损失约束以赋予模型提取判别性特征的能力。样本数量的增加能够对算法检测性能的提升起积极作用。为此,构建数据增值模块。在该模块中,利用通道随机打乱和重组的方式对自建数据集进行扩充。进一步,在细粒度信息获取模块中,采用小尺度特征学习的机制对异物进行表征。实验结果证明了该研究提出算法的优越性和有效性。瓶装矿泉水异物检测平均准确率可达96.22%,m AP值为83.84%,召回率为86.31%,检测速度为50 f/s。因此,该研究能够为瓶装矿泉水异物检测提供可靠的技术支持。

Small-scale Feature Extraction Method for Detection Foreign Matter in Bottled Mineral Water

The foreign matter detection in bottled mineral water was studied, in order to find out the foreign matter and impurities in mineral water, which has important applications in the mineral water manufacturing industry. However, the traditional machine has a low vision detection accuracy, large background interference, and high rates of missed detection and false detection. In order to solve the above problems, a small-scale feature extraction algorithm for foreign matter detection in bottled mineral water was proposed. The algorithm mainly included a discriminant feature learning module, data augmentation module, and fine-grained information acquisition module. For mineral water foreign matter, in the discriminant feature learning module, the clustering method was used to design a reasonable a prior frame size and process the feature map designing prior box size by clustering and processing the feature map. The model was given discriminant features for extraction through imposing loss constraint on the output of the network. The increase in the number of samples can have a positive effect on the improvement of the performance of detection algorithm. To achieve this, a data augmentation module. In this module, the self-built data set is expanded by random channel shuffling and reorganization. Furthermore, in the fine-grained information acquisition module, a small-scale feature learning mechanism is used to characterize foreign matter. The experimental results have proved the superiority and effectiveness of the algorithm proposed in this research. The average accuracy of foreign matter detection in bottled mineral water was 96.22%, with the mAP value as 83.84%, recall rate as 86.31%, and detection speed as 50 f/s. Therefore, this study can provide reliable technical support for detecting foreign matter in bottled mineral water.