摘要
水稻病害的多样性和复杂性使得其准确识别成为一项艰巨的任务,尤其是在非理想环境下,诸如背景噪声和病害特征提取的困难等进一步加剧了这一挑战.为了应对这些问题,该文提出了一种基于ConvNeXt网络的改进模型,即PC-ConvNeXt.该模型通过引入轻量级的金字塔切分注意力机制,有效地构建了一个多尺度特征融合模块,以更好地处理在复杂背景下的噪声问题,以及整合了通道和空间注意力机制,对特征图进行精确校准,使模型能够更准确地突出显示在图像中的病害区域.在数据增强方面,除采用传统的数据增强方法外,还结合了扩散模型来合成病害叶片图像,为模型提供了包含健康和叶片病害图像及其对应病害类别标签的综合数据集.在8种水稻病害识别数据集上的测试结果显示:PC-ConvNeXt模型展现了优异的性能,准确率和平均精度分别为88.02%和95.44%,均达到了较高水平标准.实验结果充分表明PC-ConvNeXt模型在水稻病害识别任务中的有效性和优越性.与对比模型相比,PC-ConvNeXt在精度和性能方面显著提升,展示了其在实际应用中的潜力和价值.
The identification of rice plant diseases is challenging due to their diversity and complexity,especially in non-ideal environments where issues such as background noise and difficulty in extracting disease characteristics become more pronounced.To tackle these challenges,the improved model called PC-ConvNeXt is introduced,based on the ConvNeXt network.This model incorporates a lightweight pyramid split attention mechanism to create an effective multi-scale feature fusion module,enhancing its ability to handle noise in complex backgrounds.Additionally,it integrates channel and spatial attention mechanisms to precisely calibrate feature maps,allowing the model to accurately highlight diseased regions within images.In addition to traditional methods,the study employs diffusion models are employed to synthesize images of diseased rice leaves for data augmentation,providing a comprehensive dataset including healthy and diseased leaf images with corresponding disease category labels.Tests on eight rice disease identification datasets show that the PC-ConvNeXt model exhibits outstanding performance,achieving an accuracy of 88.02%and a mean average precision of 95.44%,meeting high standards.These results validate the efficacy and superiority of the PC-ConvNeXt model in rice disease identification tasks,showing significant improvements in accuracy and performance compared to baseline models and highlighting its practical potential and value.
作者
王龙飞
李毅
曹丽萍
曹利
徐慧英
杨乐
朱信忠
谢刚
刘婷
WANG Longfei;LI Yi;CAO Liping;CAO Li;XU Huiyin;YANG Le;ZHU Xinzhong;XIE Gang;LIU Ting(School of Computer and Information Engineering,Jiangxi Agricultural University,Nanchang Jiangxi 330045,China;College of Computer Science and Technology,Zhejiang Normal University,Jinhua Zhejiang,321004,China;School of Big Data and Computer Science,Guizhou Normal University,Guiyang Guizhou 550025,China)
出处
《江西师范大学学报(自然科学版)》
北大核心
2025年第3期290-302,共13页
Journal of Jiangxi Normal University(Natural Science Edition)
基金
国家自然科学基金(61862032,62376252)
浙江省自然科学基金重点课题(LZ22F030003)
江西省教育厅科技计划课题(GJJ210432)资助项目.
