基于改进FasterNet的轻量化小麦生育期识别模型

针对现阶段小麦生育期信息获取需依靠人工观测，效率低、主观性强等问题，本文构建包含冬小麦越冬期、返青期、拔节期和抽穗期4个生育期共计4599幅小麦图像数据集，并提出一种基于FasterNet的轻量化网络模型FSST(Fast shuffle swin transformer)，开展4个关键生育期的智能识别。在FasterNet部分卷积的基础上引入Channel Shuffle机制，以提升模型计算速度。引入Swin Transformer模块来实现特征融合和自注意力机制,用来提升小麦关键生育期识别准确率。调整整个模型结构，进一步降低网络复杂度，并在训练中引入Lion优化器，加快网络模型收敛速度。在自建的数据集上进行模型验证，结果表明，FSST模型参数量仅为1.22×107，平均识别准确率、F1值和浮点运算量分别为97.22%、78.54%和3.9×108，与FasterNet、GhostNet、ShuffleNetV2和MobileNetV3 4种模型相比，FSST模型识别精度更高，运算速度更快，并且识别时间分别减少84.04%、73.74%、72.22%和77.01%。提出的FSST模型能够较好地进行小麦关键生育期识别，并且具有识别快速精准和轻量化的特点，可以为大田作物生长实时监测提供信息技术支持。

Lightweight Wheat Growth Stage Identification Model Based on Improved FasterNet

In response to the problems of low efficiency and strong subjectivity in obtaining information about the current stage of wheat development that relies on manual observation, a wheat image dataset consisting of four key growth stages of winter wheat: winterovering stage, green-turning stage, jointing stage, and heading stage, totaling 4599 images were constructed. A lightweight model FSST (fast shuffle swin transformer) based on FasterNet was proposed to carry out intelligent recognition of these four key growth stages. Firstly, based on the partial convolution of FasterNet, the Channel Shuffle mechanism was introduced to improve the computational speed of the model. Secondly, the Swin Transformer module was introduced to achieve feature fusion and self attention mechanism, it can improve the accuracy of identifying key growth stages of wheat. Then the structure of the whole model was adjusted to further reduce the network complexity, and the Lion optimizer was introduced into the training to accelerate the training speed of the model. Finally, model validation on the self-built wheat dataset with four key growth stages was performed. The results showed that the parameter quantity of the FSST model was only 1.22×107, the average recognition accuracy was 97.22%, the F1 score was 78.54%, and the FLOPs was 3.9×108. Compared with that of the FasterNet, GhostNet, ShuffleNetV2 and MobileNetV3 models, the recognition accuracy of the FSST model was higher, the operation speed was faster, and the recognition time was reduced by 84.04%, 73.74%, 72.22% and 77.01%, respectively. The FSST model proposed can effectively identify the key growth stage of wheat, and had the characteristics of fast, accurate, and lightweight recognition. It can provide a reference for optimizing the application of deep learning models in smart agriculture and offerring information technology support for real-time monitoring of field crop growth on resource-constrained mobile devices.