Research on Pear Leaf Disease Grading Based on MECB-DeepLabV3+

Chen Xiangqu; Li Aifeng; Liang Dongyue; Lyu Zhaodong; Zhang Fangheng

doi:10.14083/j.issn.1001-4942.2025.06.016

Abstract：The accurate grading of pear leaf diseases is of great significance to effective disease control and improvement of pear fruit yield and quality.Based on pear black spot disease,brown spot disease,gray spot disease and healthy leaves,this study proposed a lightweight image segmentation model for pear leaf dis-eases called MECB-DeepLabV3+.Firstly,the transfer learning method was used to compare and analyze the U-Net,PSPNet and DeepLabV3+networks,and the DeepLabV3+with the best comprehensive performance was selected as the basic network model.Secondly,to address the issues of large parameter size and high com-putational complexity in DeepLabV3+,MobileNetV3 was chosen as the backbone network to achieve model lightweighting.Finally,to overcome the shortcomings of DeepLabV3+in recognizing fine lesions and boundary segmentation,efficient channel attention(ECA),coordinate attention(CA),and bottleneck attention module(BAM)were introduced.Additionally,the Ranger21 optimizer and a composite loss function were used to op-timize model training.The experimental results showed that the mean intersection over union and mean pixel accuracy of the proposed model reached 91.22％and 95.01％,respectively,representing an improvement of 2.91 and 2.14 percentage points compared to the basic network.The parameter size of the proposed model was only 4.221 M.By using the proportion of disease spot area for disease grading,the overall average accuracy reached 95.02％,enabling effective disease grading,which could provide the scientific reference for control-ling pear leaf diseases.