关键词: 深度学习, 病害识别, ResNet, 迁移学习

Abstract:

Diseases are the main factors threatening crop yield， and their characteristics are complex and varied. It is often difficult to accurately identify them for agricultural practitioners lack professional knowledge. In the past， image recognition methods aimed at a single crop， and extracted disease features for recognition after image segmentation， which could not adapt to multiple crops. Aiming at this problem， rice， tomato， citrus and apple were as the research objects， a deep learning network framework was built based on ResNet model， and a new full connection layer containing Squeeze-and-Extension （SE） module was designed. The weight of pre-training on ImageNet was imported， and the disease model was trained on the disease data set. In order to expand the image data， the original image of the training set was increased or decreased in brightness， randomly rotated and mirror flipped. Based on the expanded training set， the disease identification and disease degree grading were studied. The results showed that the average disease degree accuracy rate of identifying rice， citrus， tomato and apple was 94.16%， and the average disease type accuracy rate was 92.45%. The disease monitoring platform was developed based on the trained model and c#.net core to realize the intelligent identification of crop diseases.

Key words: deep learning, disease recognition, ResNet, transfer learning