魏超宇 韩文 庞程 刘辉军

摘要： 由于自然環境下果蔬植株的果实、枝干和叶片等目标尺度不一、边缘不规则，因此造成其准确分割较为困难。针对该问题，提出1种多尺度特征融合和密集连接网络（Multi-scale feature fusion and dense connection networks，MDNet）以实现黄花梨疏果期植株图像的准确分割。在研究中借鉴了编码-解码网络，其中编码网络采用DenseNet对多层特征进行复用和融合，以改善信息传递方式;解码网络使用转置卷积进行上采样，结合跳层连接融合浅层细节信息与深层语义信息;在编码、解码之间加入空洞空间金字塔池化（Atrous spatial pyramid pooling，ASPP）用于提取不同感受野的特征图以融合多尺度特征，聚合上下文信息。结果表明，ASPP有效提高了模型的分割精度，MDNet在测试集上的平均局域重合度（MIoU）为77.97%，分别较SegNet、Deeplabv2和DNet提高了8.10个、5.77个和2.17个百分点，果实、枝干和叶片的像素准确率分别为93.57%、90.31%和95.43%，实现了黄花梨植株果实、枝干和叶片等目标的准确分割。在翠冠梨植株图像的独立测试中，MIoU为70.93%，表明该模型具有较强的泛化能力，对自然环境下果蔬植株图像的分割有一定的参考价值。

关键词： 黄花梨植株;多尺度特征融合;密集连接网络;图像分割;空洞空间金字塔池化（ASPP）;感受野

中图分类号： TP391.4   文献标识码： A   文章编号： 1000-4440（2021）04-0990-08

Image segmentation of Huanghua pear plants at fruit-thinning stage based on multi-scale feature fusion and dense connection network

WEI Chao-yu， HAN Wen， PANG Cheng， LIU Hui-jun

（College of Metrological Technology and Engineering， China Jiliang University， Hangzhou 310018， China）

Abstract：  As the fruits， branches and leaves of fruit and vegetable plants vary in scales and margins under natural environment， it is difficult to segment them accurately. To solve the problem， a multi-scale feature fusion and dense connection network （MDNet） was proposed to achieve the accurate segmentation of Huanghua pear images at fruit-thinning stage. The coding-decoding network was adopted in this study and DenseNet was adopted to reuse and fuse multi-layer features in the coding network， so as to improve the transfer mode of information. The transposed convolution was used to carry out up-sampling in the decoding network， and the skip-layer connection was also employed to fuse shallow detail information and deep semantic information. The atrous spatial pyramid pooling （ASPP） was added between coding and decoding to extract future maps with different receptive fields so as to fuse multi-scale feature and aggregate context information. The results showed that ASPP improved the segmentation accuracy of the MDNet model effectively. The mean intersection over union （MIoU） of the MDNet on the test set was 77.97%， improved by 8.10， 5.77 and 2.17 percentage points respectively compared with SegNet， Deeplabv2 and DNet. The pixel accuracy for fruits， branches and leaves was 93.57%， 90.31% and 95.43%， respectively. Therefore， the accurate segmentation of fruits， branches and leaves of Huanghua pear plants was realized. The MIoU was 70.93% in the independent test of Cuiguan pear plants images， indicating that the model had strong generalization ability and was particularly valuable for the image segmentation of fruit and vegetable plants in natural environment.

Key words： Huanghua pear plants;multi-scale feature fusion;dense connection network;image segmentation;atrous spatial pyramid pooling （ASPP）;receptive field

植株分割对于作物表型信息获取、生长状态监测和自动采收等具有重要意义。目前，机器视觉技术已广泛用于小麦[1]、水稻[2]和玉米[3]等大田作物的分割。果蔬多采用矮株密植，冠层结构复杂，在自然环境下，存在果叶近色、光照度不均、枝叶遮挡和尺度不一等问题，使得对其果实、枝干和叶片等部位的准确分割具有挑战性。……