基于改进ResNet18模型的饲料原料种类识别方法

Identification of Feed Raw Material Type Based on Improved ResNet18 Model

为了解决饲料生产过程中入仓原料种类采用人工取样感官识别所存在的问题,实现原料种类自动识别,以玉米、麸皮、小麦、豆粕、鱼粉等大宗饲料原料为研究对象,自主设计搭建了多通道入仓原料种类自动识别装置,采集饲料原料图像数据集,并使用数据增强的方法增加样本多样性。基于ResNet18网络模型加入通道注意力机制、增加Dropout函数,并嵌入余弦退火法的Adam优化器,引入迁移学习机制训练模型,构建适用于饲料原料种类识别的CAM-ResNet18网络模型。CAM-ResNet18网络模型的原料种类验证准确率达99.1%,识别时间为2.58 ms。与ResNet18、ResNet34、AlexNet、VGG16等网络模型相比,模型验证集准确率分别提升0.6、0.2、3.7、1.1个百分点。针对混淆矩阵结果分析,测试集识别平均准确率达99.4%,具有较高的精确度和召回率。结果表明,构建的CAM-ResNet18网络模型在饲料原料种类识别方面具有较高的识别精度和较快检测速度,自主研发的多通道入仓原料种类自动识别装置具有实际应用价值。

With the aim to solve the problem of manual sampling and sensory identification of feed raw material entering the silo in the feed production process,and realize automatic identification of raw material type,taking bulk feed raw material such as corn,bran,wheat,soybean meal and fish meal as the research object,a multi-channel automatic identification device for feed raw material type was designed and built independently,feed raw material image dataset was collected,and data augmentation methods were used to increase sample diversity.Based on ResNet18 convolution neural network,CAM-ResNet18 network model for feed raw material type identification was constructed by adding the channel attention mechanism,adding the Dropout method,adopting the Adam optimizer and embedding the cosine annealing method,while the migration learning was introduced to train the model.The average accuracy of the CAM-ResNet18 network model for feed raw material type reached 99.1% in the validation set,with a recognition time of 2.58 ms.Compared with the ResNet18,ResNet34,AlexNet and VGG16 network models,the validation accuracy was improved by 0.6,0.2,3.7 and 1.1 percentage points,respectively.For the result analysis of confusion matrix,the average accuracy of test set recognition was 99.4%,which had high accuracy and recall.The results showed that CAM-ResNet18 network model had higher accuracy rate and faster detection speed in the identification of feed raw material,providing a theoretical method and technical support for the identification of feed raw material entering the silo in the actual production.