【摘要】 为实现植物表型参数的精准获取，针对自然环境下不同尺度叶片分割的细节损失问题，提出结合空频域多尺度特征的植株叶片分割方法。以U-Net网络架构为基础，设计下采样频域变换模块，在卷积神经网络中引入频域特征表示替换池化层，利用2D-DCT和2D-IDCT的频域变换方法感知植株叶片目标的全局语义特征；构建多尺度特征融合模块，增加6个上采样节点，提取和连接植株叶片图像细粒度的特征信息；改进通道注意力模块学习分支特征，采用联合损失函数优化网络性能。在2017CVPPP公开数据集上开展实验，结果表明，植株叶片分割网络的交叉比、平均交叉比、像素准确率、精准率和F-score分数分别达到了97.07%、98.04%、99.53%、99.68%和99.74%。与FCN-8s、FCN-ResNet、DeepLabV3+、SegNet和U-Net模型相比，网络的交叉比和平均交叉比最高提升23.32%和12.43%,在较小尺度和细节处理上改善了植株叶片的分割精度，可为植物表型方向的应用研究提供一种可借鉴的思路。更多还原

【Abstract】 To improve the effect of plant segmentation and achieve accurate acquisition of plant phenotypic parameters, this study proposes a plant segmentation network to fuse spatial-frequency domain feature representation. Based on the U-Net network architecture, the frequency domain transform module is built for down-sampling, the frequency domain feature representation is introduced in the convolutional neural network to replace the pooling layer, and the frequency domain transform method of 2D-DCT and 2D-IDCT is used to perceive the global semantic features of the plant. The multi-scale feature fusion module is constructed by adding six up-sampling nodes to extract and connect the fine-grained feature information of the plant image. The channel attention module is modified to learn branch features and the hybrid loss function is employed to optimize the network. Experiments are conducted on the 2017 CVPPP public dataset, and the results show that the intersection over union, mean intersection over union, pixel accuracy, precision and F1 of the plant segmentation network reach 97.07%, 98.04%, 99.53%, 99.68% and 99.74%, respectively. Compared with the FCN-8s, FCN-ResNet, DeepLabV3+, SegNet and U-Net models, the intersection over union and mean intersection over union of the network were improved by up to 23.32% and 12.43%. The proposed method can improve the segmentation accuracy of plant at smaller scales in detail processing, and it can provide a useful idea for applied research in the plant phenotype.更多还原