基于改进U-Net网络的苹果叶部病害语义分割方法

Semantic segmentation method of apple leaf disease based on improved U-Net network

针对自然条件下苹果叶部病斑分割与识别效果欠佳的问题,本文提出一种融合条件随机场和卷积块状注意力模块的苹果叶部病害语义分割模型,实现苹果叶部锈病、褐斑病、灰斑病及斑点落叶病的病斑准确分割和识别。本文在U-Net模型基础上,使用ResNet50为骨干网络防止梯度消失问题,并分别在跳跃连接分支与上采样层加入卷积块状注意力模块,减少训练过程中的分割精度损失,融合Dice Loss和Focal Loss降低损失波动,最后利用条件随机场优化分割结果,获取病斑掩模图像,实现对苹果叶部病害语义分割。本研究在自制苹果叶部病害数据集上进行试验,分析了光照、阴影及水滴等因素对分割结果的影响。试验结果表明:本文构建的语义分割模型相比传统U-Net模型,平均分割精度(mIoU)提升8.24百分点,平均分类精度(mPrecision)提升11百分点,类别平均像素准确率(mPA)提升6.09百分点,受光照不均、雨滴的影响更小,具有更好的鲁棒性和可靠性。

Aiming at the problem of poor segmentation and recognition of apple leaf spots under natural conditions,this paper proposed a semantic segmentation model for apple leaf diseases that incorporated conditional random fields and convolutional block attention modules to achieve accurate segmentation and recognition of spots of apple leaf rust,brown spot,gray spot and Alternaria leaf spot disease.In this paper,based on the U-Net model,ResNet50 was used as the backbone network to prevent the gradient vanishing problem,and the convolutional block attention module was added to the jump-connected branch and the up-sampling layer respectively,to reduce the loss of segmentation accuracy during the training process,and the fusion of dice loss and the cross-entropy loss function to reduce the loss fluctuation,and finally,the segmentation results were optimized using the conditional random field to obtain the diseased spot mask image,which was used to realize the accurate segmentation and recognition of apple leaf rust,brown spot,grey spot and Alternaria leaf spot disease so as to realize semantic segmentation of apple leaf diseases.In this study,we conducted experiments on the homemade apple leaf disease dataset,and analyzed the effects of light,shadow and water droplets on the segmentation results.The experimental results showed that the semantic segmentation model constructed in this paper improved the average segmentation accuracy mIoU by 8.24 percentage points,the average classification accuracy mPrecision by 11 percentage points,and the average pixel accuracy of category mPA by 6.09 percentage points compared with the traditional U-Net model,and was less affected by uneven illumination and raindrops,and had better robustness and reliability.