中国农业大学学报2026，Vol.31Issue(5)：207-222,16.DOI:10.11841/j.issn.1007-4333.2026.05.17
基于改进DeepLabV3+的玉米杂草密度提取方法研究
Research on maize weed density extraction method based on improved DeepLabV3+
摘要
Abstract
To address the issues of variable visual features of maize weeds in complex field environments in hilly and mountainous areas,insufficient accuracy of traditional segmentation models,and difficulty in balancing lightweight design and real-time performance,this study proposes an improved DeepLabV3+semantic segmentation model for high-precision weed identification and density extraction.Methodologically,the lightweight MobileNetV2 network replaces the original Xception backbone.A SEMA-ASPP module integrating a channel attention mechanism(Squeeze-and-Excitation Networks,SE-Net)and an adaptive activation function(Meta-ACONC)is designed to replace the ReLU function in the original Atrous Spatial Pyramid Pooling module,aiming to enhance multi-scale feature extraction capability and generalization performance.An Efficient Local Attention(ELA)mechanism is introduced into the decoder to optimize weed edge segmentation.Meanwhile,a combined strategy of cross-entropy loss and Dice coefficient loss is adopted to effectively address the imbalance between target and background regions in the dataset.Experimental results show that:1)While maintaining high segmentation accuracy,the improved model significantly enhances computational efficiency.Its mean intersection over Union(mloU)and mean Pixel Accuracy(mPA)reach 92.72％and 95.05％,respectively,representing improvements of 3.22 and 2.95 percentage points over the original DeepLabV3+model.2)The model's computational cost and parameter count are only 51.84 GFLOPS and 5.89x106,accounting for 31.04％and 10.77％of the original model,respectively,with an inference speed reaching 120.51 frames per second on GPU.3)Compared with mainstream models including DeepLabV3+,SegFormer,PSPNet,U-Net and HRNet,the proposed model achieves lower computational cost and faster inference speed while ensuring comparable or high segmentation accuracy.Based on the segmentation results,a sliding window scanning algorithm is further employed to extract weed density.The coefficient of determination R2 for the linear regression between the predicted and actual density reaches 0.981,which verifing the reliability of the model.The proposed model achieves a favorable balance among accuracy,efficiency and lightweighting design,providing technical support for variable-rate spraying in precision agriculture.关键词
玉米杂草/密度提取/语义分割/DeepLabV3+/轻量化模型/注意力机制Key words
maize weed/density extraction/semantic segmentation/DeepLabV3+/lightweight model/attention mechanism分类
农业科技引用本文复制引用
安美林,赵建国,赵学观,王雅雅,马志凯,李媛普,王博奥,郝建军..基于改进DeepLabV3+的玉米杂草密度提取方法研究[J].中国农业大学学报,2026,31(5):207-222,16.基金项目
国家重点研发计划(2023YFD2301500) （2023YFD2301500）
河北省现代农业产业技术体系创新团队建设项目(HBCT2024030207) （HBCT2024030207）
