【作者】 张智；

【导师】 刘洋洋； 鲍恩财；

【作者基本信息】 安徽农业大学 ， 农业硕士（专业学位）， 2024， 硕士

【摘要】 我国作为葡萄的生产大国,占全球产量的56.58%,但其生产深受病害影响,各类病害每年都会给葡萄产业造成巨大的经济损失。因此,利用计算机视觉技术对葡萄叶片病害进行预防性诊断显得尤为重要。目前,农户主要通过人工检测和定期喷洒农药的方式来防治葡萄病害,但该方法不仅识别效率低下,而且农药使用量大,违背了绿色环保的可持续发展理念。随着图像识别技术的进步,许多学者已经提出了基于深度学习的图像识别技术。但深度学习技术对计算硬件的要求高,训练时间长。本文基于宽度学习系统和深度学习技术,提出一种葡萄叶片病害识别与病情程度检测方法,能实现快速训练、及时准确的病害识别,并针对不同阶段的病害提供相应的防治建议。本文的主要研究内容包括:(1)基于宽度学习系统的自适应性葡萄叶片病害识别方法。本研究针对Adam算法进行优化,提高其学习率的稳定性,并代替宽度学习中的岭回归运算,同时在宽度学习的映射层和增强层之间增加SENet注意力机制,构建出Ada LTM-BLS-SE网络模型,通过实验验证该模型对葡萄叶片病害的识别精准度达95.32%,单张图像平均检测时间为0.145 s,证明模型在保证快速训练的同时有较高的识别准确度。(2)基于U-Net网络的病斑分割算法优化。本研究基于U-Net分割模型,集成空洞空间金字塔池化和特征金字塔网络,利用不同膨胀率的膨胀卷积捕捉多尺度上下文信息,增强了对图像细节和全局结构的理解,同时在解码过程中提供丰富的语义信息和高分辨率细节,提高了模型对细小病斑的识别能力。经测试MSCU-Net分割模型的MIOU达到86.61%较原始U-Net模型提高6.48%,MPA为90.23%,较原始U-Net模型提高8.09%。(3)基于层次分析法制定病害等级。本研究采用层次分析法,针对不同病害类型和病斑面积,制定病害等级,为精准施药技术提供指导,实现精细化施药,提高农药使用效率,增强防治效果,为果园绿色植保技术的发展提供了理论与技术支持。更多还原

【Abstract】 As a major producer of grapes,China accounts for 56.58 % of global production,but its production is deeply affected by diseases.Various diseases cause huge economic losses to the grape industry every year.Therefore,it is particularly important to use computer vision technology to carry out preventive diagnosis of grape leaf diseases.At present,farmers mainly control grape diseases by manual detection and regular spraying of pesticides.However,this method is not only inefficient in identification,but also uses a large amount of pesticides,which violates the concept of sustainable development of green environmental protection.With the progress of image recognition technology,many scholars have proposed image recognition technology based on deep learning.However,deep learning technology has high requirements for computing hardware and long training time.Based on the broad learning system and deep learning technology,this paper proposes a method for grape leaf disease identification and disease degree detection,which can achieve rapid training,timely and accurate disease identification,and provide corresponding prevention and control suggestions for different stages of disease.The main research contents of this thesis include:(1)Adaptive grape leaf disease recognition method based on Broad Learning System:This study optimizes the Adam algorithm to improve its learning rate stability,replacing ridge regression operations in BLS.Additionally,the SENet attention mechanism is integrated between the mapping layer and the enhancement layer of BLS,constructing the Ada LTM-BLS-SE network model.Experimental results validate that the model achieves a recognition accuracy of 95.32% for grape leaf diseases,with an average detection time of0.145 seconds per image,demonstrating high recognition accuracy while ensuring rapid training.(2)Optimization of lesion segmentation algorithm based on U-Net network: This study enhances the U-Net segmentation model by integrating atrous spatial pyramid pooling and feature pyramid networks,utilizing atrous convolution with different dilation rates to capture multi-scale contextual information.This improves the understanding of image details and global structures,while providing rich semantic information and high-resolution details during the decoding process.The MSCU-Net segmentation model achieved a Mean Intersection over Union(MIOU)of 86.61%,an improvement of 6.48%over the original U-Net model,and a Mean Pixel Accuracy(MPA)of 90.23%,an improvement of 8.09% over the original U-Net model.(3)Based on the analytic hierarchy process,the disease grade is determined.In this study,the analytic hierarchy process was used to formulate the disease grade for different disease types and lesion areas,so as to provide guidance for precise application technology,realize fine application,improve the efficiency of pesticide use,enhance the control effect,and provide theoretical and technical support for the development of green plant protection technology in orchards.更多还原