【作者】 李赟；

【导师】 张继军； 张茜；

【作者基本信息】 山东农业大学 ， 农业工程与信息技术（专业学位）， 2023， 硕士

【摘要】 中国是全球最大的樱桃生产国之一,樱桃种植产业已成为人民的支柱性产业。樱桃一旦发生病害,不仅影响樱桃当年产量和品质,还会影响来年花芽质量,进而造成种植户收入减少。传统的深度学习病害识别模型虽然精度高但内存需求和计算量较大,对软硬件设备要求较高,无法适用于种植户常用的移动设备或嵌入设备,因此构建轻量化的樱桃病害识别模型并应用于移动端设备尤为重要。本文将樱桃叶部病害图像作为研究对象,重点研究樱桃病害识别的轻量化模型设计和应用。以Mobile Net V1模型为基础,采用自制的樱桃叶部图像数据集,引用通道剪枝和注意力机制的模型压缩算法,构建了樱桃叶部病害识别轻量化模型PSM-Net。主要研究成果如下:(1)针对Mobile Net V1模型在樱桃病害数据集中可能存在的过拟合问题,提出了改进的迁移学习的模型训练方式。首先以健康叶片和细菌穿孔病、褐斑病、白粉病、缩叶病5种叶片类型制作樱桃病害图像数据集,采用传统数据处理技术,包括裁剪、翻转、添加噪声等方法,对数据进行增强并扩充数据集数量。其次设计迁移学习训练方式,将在源数据域Plant village上训练的Mobile Net V1模型迁移到目标域自制樱桃病害数据集上,微调Mobile Net V1模型的顶层结构将分类数量修改为5。经对比实验表明,基于迁移学习的Mobile Net V1模型识别准确率达到91.48%,比直接训练模型提高了5.53%,且在新的数据集上未表现出过拟合现象,收敛速度显著加快。(2)针对模型在移动端部署时占用内存过大计算量大等问题,构建了樱桃叶部病害识别轻量化模型PSM-Net。首先在Mobile Net V1模型基本单元逐点卷积的BN层中引入缩放因子,按照缩放因子的大小设置通道重要性排序。其次在Mobile Net V1模型下采样单元中引入改进的SE注意力机制,加强特征提取能力。最后用SELU替换Re LU激活函数,提升模型推理速度。设计对比实验表明,剪枝比例80%得到的PSM-Net模型,模型尺寸为16.87 M,比原模型减小了62%;参数量为1.18 M,比原模型减小66.3%;推理时间为17ms,比原模型加快67.3%,识别准确率为90.98%,仅损失0.5%。本研究设计的PSM-Net模型,其参数大小和模型尺寸等均优于其他轻量级神经网络,可适用于移动端设备。(3)构建樱桃叶部病害识别系统。使用Java语言设计开发网页端后台管理系统和移动端病害识别系统,实现了樱桃叶部病害图像的实时检测。更多还原

【Abstract】 China is one of the world’s largest cherry producers,and the cherry planting industry has become a pillar industry for the people.Once the cherry disease occurs,it will not only affect the yield and quality of the cherry in the current year,but also affect the quality of the flower bud in the next year,which will cause the income of farmers to decrease.Although the traditional deep learning disease recognition model has high accuracy,it has large memory and computational requirements,and has high requirements on hardware and software devices,which cannot be applied to mobile devices commonly used by growers or embedded devices.Therefore,it is particularly important to build a lightweight cherry disease recognition model and apply it to mobile devices.This paper focuses on the design and application of lightweight model for cherry disease recognition,taking the image of cherry disease as the research object.Based on the Mobile Net V1 model,a lightweight cherry leaf disease recognition model PSM-Net was constructed by using a self-made cherry leaf image dataset and the model compression algorithm of channel pruning and attention mechanism.The main research results are as follows:(1)Aiming at the possible overfitting problem of Mobile Net V1 model in cherry disease data set,an improved transfer learning model training method was proposed.Firstly,cherry disease image data sets were made from five leaf types: healthy leaves,bacterial perforation disease,brown spot disease,powdery mildew disease and leaf shrinkage disease.Traditional data processing techniques,including clipping,flipping and adding noise,were used to enhance the data and expand the number of data sets.Secondly,the transfer learning training method was designed.The weight of the Mobile Net V1 model trained on the source data domain Plant village was transferred to the self-made cherry disease data set in the target domain,and the top-level structure of the Mobile Net V1 model was fine-tuned to modify the number of categories to 5.Comparative experiments show that the recognition accuracy of the Mobile Net V1 model based on transfer learning reaches 91.48%,which is 1.43% higher than that of the direct training model.Moreover,there is no fitting phenomenon on the new data set,and the convergence speed is significantly accelerated.(2)A lightweight model PSM-Net for cherry leaf disease identification was constructed to address issues such as excessive memory usage and computational complexity when deploying models on mobile devices.Firstly,a scaling factor is introduced into the BN layer of the pointwise convolution of the basic units in the Mobile Net V1 model,and channel importance sorting is set according to the size of the scaling factor.Secondly,an improved SE attention mechanism is introduced into the downsampling unit of the Mobile Net V1 model to enhance feature extraction capability.Finally,replace the Re LU activation function with SELU to improve the inference speed of the model.The design comparison experiment shows that the PSM Net model obtained with a pruning ratio of 80% has a model size of 16.87 M,which is62% smaller than the original model;The parameter quantity is 1.18 M,which is 66.3% less than the original model;The inference time is 17 ms,which is 67.3% faster than the original model,and the recognition accuracy is 90.98%,with only a loss of 0.5%.The PSM Net model designed in this study has better parameter size and model size than other lightweight neural networks,and can be applied to mobile devices.(3)The identification system of cherry leaf disease was constructed.Java language was used to design and develop the webpage back-end management system and the mobile terminal disease identification system,and realized the real-time detection of cherry leaf disease image更多还原