基于SSD的马铃薯叶片图像病斑分割算法研究

Research on Leaf Image Disease Spot Segmentation Algorithm Based on SSD Network

摘要: 为解决传统特征分割对光照不匀、噪声差异图像的叶片病斑提取适应性问题，提出一种基于SSD神经网络的深度学习方法。先通过多尺度卷积提取深度特征，产生病斑边界候选框集合，再利用非极大值抑制获取病斑叶片区域。之后对于获得的病斑区域利用SVM进行像素分割得到病斑精准边界。对马铃薯叶片图像进行实验验证，获得了图像中病斑的准确边界，病斑区域定位准确率达85%以上，在该基础之上的边界分割准确率可达到80%。相比其他病斑分割方法，该方法利用化整体分割为局部分割的思想，实现了复杂背景下的马铃薯病斑的有效分割，为下一步马铃薯病斑识别打下基础。

Abstract: Image-based automatic segmentation of leaf disease spot is the basis for plant disease identification. A deep learning method based on SSD neural network is proposed in order to solve the problem of adaptive extraction of disease spot from uneven image and noise difference images in traditional feature segmentation. The idea of the algorithm is to transform the whole segmentation into partial segmentation. The depth features are extracted by multi-scale convolution to generate a set of leaf disease spot bounding-boxes, and the disease spot regions of leaf are obtained by non-maximum suppression. Experiments on the image of potato leaves were carried out to obtain the exact boundary of the disease spots in the image. The location accuracy of the disease spot was more than 85%, and the accuracy of boundary segmentation was 80%. Compared with the traditional method of segmentation of disease spot, this method proposes an effective solution for complex background and automatic segmentation of multiple disease spots, which lays a foundation for further leaf disease identification.