摘要
针对复杂背景和多目标遮挡导致的检测精度下降问题,提出了一种基于深度学习的轻量化茶叶病虫害检测方法。该方法在现有DETR模型基础上引入小波变换-卷积模块,在减少模型参数量的同时显著提升了对多尺度特征的捕获能力;结合多尺度多头注意力机制,实现了跨尺度的全局特征融合,有效克服了传统注意力机制在小目标检测中的局限性;通过设计上下文引导空间特征重建特征金字塔网络,进一步提升复杂场景下目标检测的鲁棒性和精确性。实验结果表明,模型识别准确率达97.7%,参数量和浮点运算量均降低35%以上;通过在树莓派平台部署验证,表明所提方法能够准确、高效地完成茶叶病虫害检测任务。
A lightweight tea disease and pest detection method based on deep learning is proposed to address the issue of decreased detection accuracy caused by complex backgrounds and multiple target occlusions.Firstly,building on the existing DETR model,the incorporation of wavelet transform convolution notably amplifies the ability to capture multi-scale features,while diminishing the model's parameter count.Secondly,combined with the multi-scale and multi-head attention mechanism,the proposed method achieves cross-scale global feature fusion,effectively overcoming the shortcomings of traditional attention mechanisms in small target detection.Thirdly,a context-guided spatial feature reconstruction feature pyramid network is employed,which further improves the robustness and accuracy of object detection in complex scenes.By conducting data analysis and performance evaluation,the identification accuracy rate achieves 97.7%,the parameter count and floating-point operation reduce by over 35%.Finally,the experiments deployed on the Raspberry Pi platform show that the proposed method can accurately and efficiently detect tea leave diseases and pests.
作者
宋军
张佑丞
徐锋
焦万果
SONG Jun;ZHANG Youcheng;XU Feng;JIAO Wanguo(College of Information Science and Technology/Artificial Intelligence,Nanjing Forestry University,Nanjing 210037,China)
出处
《实验室研究与探索》
北大核心
2025年第8期39-47,54,共10页
Research and Exploration In Laboratory
基金
江苏省本科高校“理工类公共基础课程教学改革研究”专项课题(2024LGJK034)。
