摘要
针对现有深度学习模型在复杂背景下虫害特征提取能力差、泛化性能低等问题,提出一种基于改进脱粒网模型的农作物虫害识别方法。引入批量通道归一化模块,提高模型的泛化能力;将高层筛选特征金字塔网络与自定义卷积模块融合,形成多尺度特征融合模块,把模块嵌入密集连接的模型与谐波之间,增强模型特征提取能力;调整模型整体架构,得到驱虫脱粒网(TNP)模型。通过自建数据集P28对比实验,结果表明,与改进前模型相比,TNP模型的准确率提高了3.95%、参数量下降了6.47M、浮点运算量(FLOPs)下降了0.23G;与ResNet50、DenseNet、EfficientNet B4等模型相比,TNP模型的准确率、参数量、FLOPs和推理时间均有较好表现。TNP模型能够快速、准确识别农作物虫害特征信息,为及时防治虫害提供技术支持。
Aiming at the problems of poor feature extraction ability and low generalization performance of existing deep learning models for pests in complex backgrounds,a crop pest recognition method based on improved ThreshNet model is proposed.Firstly,a bulk channel normalization module(BCN)is introduced to improve the generalization ability of the model;secondly,a high level screening feature pyramid network(HSFPN)is fused with a customized convolution module(partial-conv)to form a multi-scale feature fusion module(PCHS).Secondly,the module is embedded into the model between the dense connections and the harmonic dense connections to enhance the model's feature extraction capability.Finally,the overall architecture of the model is adjusted to obtain the improved TNP model.Comparison experiments are carried out on the self-built dataset P28.The experimental results show that compared with the pre-improvement model,the accuracy of the TNP model increases by 3.95%,the number of parameters decreases by 6.47M,and the amount of floating-point operations(FLOPs)decreases by 0.23G.Compared withResNet50,DenseNet,EfficientNet B4 and other models,the TNP model has better performance in terms of accuracy,parameter quantity,FLOPs and inference time.The improved model can quickly and accurately identify the characteristic information of crop pests,and provide technical support for timely prevention and control of pests.
作者
任喜伟
孙悦
杨虹
何立风
REN Xiwei;SUN Yue;YANG Hong;HE Lifeng(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi’an 710021,China;Xi’an Friendship Medical Electronics Co.,Ltd.,Xi’an 710075,China)
出处
《实验室研究与探索》
CAS
北大核心
2024年第12期79-85,107,共8页
Research and Exploration In Laboratory
作者简介
任喜伟(1981-),男,陕西西安人,博士,教授级高工,研究方向为数字图像处理、智慧教育、过程系统工程。Tel.:13659181420,E-mail:renxiwei@sust.edu.cn;通信作者:何立风(1963-),男,湖南永州人,博士,教授,研究方向为图像处理、模式识别。Tel.:13474061776,E-mail:helifeng@istaichi-pua.jp。
