基于改进DeepLabV3+的苹果叶面病斑语义分割方法

打开文本图片集

中图分类号：TP391.41；S661.1 文献标识码：A 文章编号：2095-5553（2025）08-0075-08

Abstract：Taking appleleaf lesionsas the research object，a semantic segmentation method based onan improved DeepLabV3+model for apple leaf disease spots is proposed to solvetheproblems of low segmentation acuracy and large numberof model parameters.In this paper，MobileNetV2 isemployedas the backbone featureextraction network of DeepLabV 3+ to reduce the model's parameter size. An MP—DenseASPP module is introduced，which adds the dilated convolution layers on topof ASPP and incorporates dense connections，along with a mixed poling module to enlarge the model'sreceptive fieldandenhance itsrobustness.Amulti-scaleshalow feature layerisdesigned toimprove the segmentation capability for multi-scaletargets.Furthermore，an improved AFFmodule，named ECAFF，is proposed to merge features across diferent levelsof themulti-scaleshallow feature layer，thereby boosting inter-layer feature fusion. The results show that the mean Intersection over Union（mIoU），mean pixel accuracy and F1 score of the improved DeepLabV3+ model on the ATLDSD dataset reach 72.22% ， 88.77% and 83.44% ，representing increases of 1.10% ， 4.73% and 1.02% ，respectively，compared to the original model. The floating-point operations and parameter number of the improved model are approximately reduced by 58.5% and 77.1% ，respectively，the average frame rate is increased by 6.67f/s compared to the original model. The proposed method significantly reduces the model's parameter number whileensuring theaccuracyofleaf diseasespot detection，meting thereal-timerequirementsandlayingafoundation for online detection of leaf disease spots based on semantic segmentation.

Keywords： apple leaves； disease segmentation; atention mechanism； feature fusion; deep learning

0 引言

苹果是我国产量最大的水果，而病害是影响苹果产量的重要因素之一[1]。（剩余12098字）