Abstract: Accurate identification of silkworm day-age contributes to precise feeding and animal welfare. In order to accurately identify 14 day-age from the first day of age 3 to the seventh day of age 5 in the growth period of silkworm, this paper constructs a dataset with 14 day-age units by collecting specific silkworm species in a real environment. This paper proposes Moga-ResNet method based on an improved residual neural network. This method introduces a multi-order gating mechanism to obtain the saliency features of day-age images based on the classical residual neural network ResNet50. Through model training and testing on the same domestic silkworm day-age dataset, the recognition accuracy of Moga-ResNet is 96.57%, the F1 value is 96.57%, and the recall rate is 96.62%. Compared with the evaluation indexes of four classic models, including Swin Transformer, MobileNet v3, CSPNet and DenseNet, Moga-ResNe achieved a stronger recognition ability in day-age recognition of domestic silkworm, which can provide a foundation for carrying out work related to precise feeding and digital management.

Key words: silkworm, day-age identification, multi-order gating mechanism, residual neural network

摘要： 家蚕日龄的准确识别有助于精准饲喂和动物福利,因此为准确识别家蚕生长时期中3龄第1天至5龄第7天,共14个日龄,在实际环境下采集特定家蚕品种,构建以14个日龄为单位的数据集。提出一种基于改进残差神经网络的Moga-ResNet,该方法在经典残差神经网络ResNet50的基础上,引入多阶门控机制以获取日龄图像的显著性特征。通过在同一个家蚕日龄数据集上开展模型训练与测试得到,Moga-ResNet的识别准确率为96.57%,F1值为96.57%,召回率为96.62%,与Swin Transformer、MobileNet v3、CSPNet和DenseNet四个经典模型的评价指标相比,Moga-ResNet在家蚕的日龄识别中具有较强的识别能力,可以为开展家蚕精准饲喂和数字化管理相关工作提供基础。

关键词: 家蚕, 日龄识别, 多阶门控机制, 残差神经网络