多特征融合的轻量级马铃薯分类研究

Lightweight potato classification method based on multi-feature fusion

针对复杂环境下田间马铃薯品种识别精度不高,模型体积较大等问题,提出了一种基于多特征融合的轻量级卷积神经网络LRnet。LRnet网络将ShuffleNetV2和MobileNetV2裁剪后作为双分支网络结构的主干,分别提取马铃薯叶片和马铃薯块茎的图像特征,然后通过提出的融合策略将特征进行融合,从而对马铃薯的29个品种进行分类。实验结果表明,本文提出的LRnet相比ShuffleNetV2、MobileNetV2和ResNet不仅显著提高了马铃薯品种的分类精确度,而且模型小,易于端侧部署,同时也表明了多部位特征融合方法可以显著提高物种的识别精度,为马铃薯的品种鉴定提供了技术参考,其融合策略也可为相关研究人员提供一定的研究思路。

Aiming at the problems of low accuracy and large model size of interfield potato varieties in complex environment,a lightweight convolutional neural network(LRnet)based on multi-feature fusion was proposed.LRnet network cuts ShuffleNetV2 and MobileNetV2 as the backbone of the two-branch network structure,extracts the image features of potato leaves and potato tubers respectively,and then fuses the features through the proposed fusion strategy to classify the 29 varieties of potato.Experimental results show that compared with ShuffleNetV2,MobileNetV2 and ResNet using potato leaves for classification,the proposed LRnet not only significantly improves the classification accuracy of potato varieties,but also has a small model and is easy to be deployed end-to-end.This method also shows that the multi-part feature fusion method can significantly improve the accuracy of species identification,which provides technical reference for potato variety identification,and its fusion strategy can also provide certain research ideas for relevant researchers.