基于DeepLab v3+的葡萄叶片分割算法

为解决自然光照环境下复杂背景葡萄叶片图像的自动分割问题,使用一种DeepLab v3+语义分割算法,完成对葡萄叶片分割。该算法采用ResNet 101作为主干网络进行特征抽取;采用空洞卷积和编码模块进行多尺度特征融合,将ResNet的中间信息和编码模块的特征组合作为解码输入;采用上采样的方式进行尺度还原,得到语义分割结果。采用Cityscapes的预训练模型,将300张不同环境下、不同类型的葡萄叶片照片作为训练集,以数据增强的方式进行数据扩容,提高模型的鲁棒性和泛化能力。试验结果证明:本方法有较好的分割效果,在数据增强的方式下精确度(ACC)平均值为98.6%,较全卷积神经网络提高7.3%。对不同类型葡萄叶片分割精确度(ACC)值均高于97%,最高可达98.8%,平均交并比(mIOU)值均高于94%,最高可达97.1%。本算法能够较精准地分割自然光照条件下的葡萄叶片图像,可为后续的病害检测和病斑提取提供参考。

Grape Leaf Segmentation Based on DeepLab v3+

This research aims to solve the problem of automatic segmentation of grape leaf image with complex background under natural light environment.This research uses a DeepLab v3+semantic segmentation algorithm to complete the segmentation of grape leaves.In this algorithm,ResNet 101 is used as the backbone network for feature extraction.dilated convolution and coding module are used for multi-scale feature fusion,and the combination of the intermediate information of ResNet and the features of coding module is used as the decoding input.The upper sampling method is used for scale reduction to obtain the semantic segmentation results.By using cityscape's pre training model,300 photos of grape leaves in different environments and different types are taken as training sets,and data capacity is expanded in the way of data enhancement to improve the robustness and generalization ability of the model.The experimental results show that this method has a good segmentation effect,the average accuracy(ACC)is 98.6% in the way of data enhancement,which is 7.3% higher than that of full convolution network.The ACC value of different types of grape leaves was higher than 97%,the highest value was 98.8%,and the average IOU value was higher than 94%,the highest value was 97.1%.This algorithm can segment the grape leaf image accurately under the natural light condition,and provide the basis for the subsequent disease detection and disease spot extraction.