关键词: 植物表型, 智能检测, 深度学习, 实例分割, 迁移学习, 特征融合

Abstract:  For the intelligent detection problem of highthroughput automatic acquisition of plant phenotype images, an intelligent plant phenotype detection and segmentation algorithm was designed using transfer learning and improved Mask Region Convolution Neural Network (Mask R-CNN). Firstly, the Residual Network (ResNet) was optimized, and the Feature Pyramid Network (FPN) was used to extract the features of the input image. Then, the lengthtowidth ratio and threshold of the anchor box in the network (RPN) of candidate area was adjusted, and the spatial information of the feature map was retained by bilinear interpolation in RoIAlign. Finally, the mask detection head was improved and the feature fusion mechanism was added to obtain highquality masks. The character phenotype data set of watermelon mutant growth was trained and detected, and the results showed that the improved Mask R-CNN showed better detection performance. Compared with the traditional Mask R-CNN, detection accuracy was improved by 22%, mask accuracy was improved by 2.7%, and detection time was reduced by 42 ms, which provided technical support for improving the level of agricultural precision and promoting the development of intelligent agriculture.

Key words:  plant phenotype, intelligent detection, deep learning, case segmentation, transfer learning, feature fusion