基于轻量化模型的杂草识别方法研究毕业论文.docxVIP

基于轻量化模型的杂草识别方法研究 摘　要：智能农业正在飞速发展，而农业的绿色发展则是必然的，这就要求我们要充分运用现代化的科技手段，来实现农业生产的节本增效。我国是一个农业大国，拥有全球第一的粮食生产和全球第三的耕地，但是，我国的耕地质量却很差，其中优质耕地不足3%，因此，实现农业的可持续发展就显得尤为重要。随着杂草的日益严重，我国每年都要施用数百万吨以上的化学除草剂，其过量的长期使用造成了严重的环境污染和食品安全问题。本论文采用深度学习算法，开展了农田中杂草靶标检测技术的研究。对于杂草的分类研究，模型的训练和分类的一般是在服务器端进行的。但是这种方式存在一些问题，比如计算资源量大，速度较慢，而且准确率可能不高。以MobileNetV2和ResNet50为基础训练一个轻量级模型。在训练过程中，为了优化移动端模型，通过给Softmax损失函数加上权重和使用高精度的服务器端模型来进行指导，最终得到了一个性能较高且占用资源较少的轻量级模型。 关键词：杂草识别；深度学习；轻量化模型；损失函数；知识蒸馏 RESEARCH ON WEED IDENTIFICATION METHOD BASED ON LIGHTWEIGHT MODEL Abstract: Intelligent agriculture is rapidly developing, and the green development of agriculture is inevitable. This requires us to fully utilize modern technological means to achieve cost saving and efficiency increase in agricultural production. China is an agricultural country with the worlds largest food production and the worlds third largest arable land. However, the quality of arable land in China is very poor, with less than 3% of high-quality arable land. Therefore, achieving sustainable agricultural development is particularly important. With the increasing severity of weeds, China applies millions of tons or more of chemical herbicides every year, and their excessive long-term use has caused serious environmental pollution and food safety issues. This paper adopts deep learning algorithms to conduct research on weed target detection technology in farmland. For weed classification research, model training and classification are generally conducted on the server side. However, this method has some problems, such as high computational resources, slow speed, and possibly low accuracy. Train a lightweight model based on MobileNetV2 and ResNet50. In the training process, in order to optimize the mobile terminal model, the Softmax loss function was weighted and the high-precision server model was used to guide, and finally a lightweight model with high performance and less resource consumption was obtained. Keywords: weed identification; deep learni