基于深度卷积网络的葡萄簇检测与分割

Grape cluster detection and segmentation based on deep convolutional network

[目的]在果园场景下,簇粘连、杂物遮挡给高精度葡萄簇检测与分割造成很大难题。[方法]该文以真实种植场景下的葡萄簇为研究对象,以相机拍摄图像为数据源,提出基于2大骨干网络R50、R101与2种任务网络Mask RCNN、Cascade Mask R-CNN交叉结合的多种葡萄簇检测与分割并行化模型。对5个品种137张共2020个实例标注葡萄簇个体进行研究,为丰富数据集、提升模型泛化能力,对原始数据集随机进行改变亮度、加入高斯噪声及翻转180°操作,共获得标注图片685张。为探究不同骨干网络对模型检测与分割的影响状况,选取R50与R101对输入图像分别进行特征提取,并在Mask R-CNN和Cascade Mask R-CNN两大任务网络上进行试验。[结果]对于检测任务,Mask R-CNN-R50在AP0.75指标上比Mask R-CNN-R101提升22.3%;对于分割任务,Cascade Mask R-CNNR50在各AP指标上比Cascade Mask R-CNN-R101提升2%~13.5%。为验证学习率超参数对预测结果影响,选用6个不同学习率在Mask R-CNN-R50与Cascade Mask R-CNN-R50模型上进行试验,结果表明,随着学习率的增加,检测与分割各AP指标均先增加后减小;为探究模型的鲁棒性,将测试集图片分为深度分离、浆果粘连、杂物遮挡3大类并进行可视化分析,结果表明,Cascade Mask R-CNN-R50模型在3种场景下分割与检测效果最佳,Mask RCNN-R101效果最差。[结论]综合分析,本文Cascade Mask R-CNN-R50模型可更为精确、有效地对不同种植场景葡萄簇进行分割与检测,其可为后续葡萄自动化采摘提供模型支撑。

[Objective]In the orchard scene,cluster adhesion and debris occlusion pose a great problem for high-precision grape cluster detection and segmentation.[Methods]In present study,grape clusters in the real planting scenes were used as the research object and the images taken with the camera as the data source.A parallel model of grape cluster detection and segmentation was constructed based on the cross combination of two backbone networks(R50 and R101)and two task networks(Mask R-CNN and Cascade Mask R-CNN).A total of 137 images from 5 grape varieties of 2020 labeled grape clusters were analyzed.In order to enrich the data set and improve the generalization ability of the model,the original data set was randomly changed in brightness,Gaussian noise and 180 flipped operation on the original data set,and a total of 685 images were obtained.In order to explore the impact of different backbone networks on model detection and segmentation,Res Net50(R50)and Res Net101(R101)were selected for feature extraction of the input images,and the verification was performed on two task networks,Mask R-CNN and Cascade Mask R-CNN.[Results]The experiment showed that Mask R-CNN-R50 was 22.3%higher than Mask R-CNN-R101 in AP0.75 indicators for detection tasks;For segmentation tasks,Cascade Mask R-CNN-R50 was 2%~13.5%higher than Cascade Mask R-CNN-R101 in AP indicators.In order to verify the influence of the learning rate hyperparameters on the prediction results,six different learning rates were selected to test on the Mask R-CNN-R50 and Cascade Mask R-CNN-R50 models.The results showed that as the learning rate increases,the AP indicators for detection and segmentation were increased first and then decreased;In order to explore the robustness of the model,the test set pictures were divided into three categories:deep separation,berry adhesion,and debris blocking and visual analysis was performed.The results indicated that the Cascade Mask R-CNN-R50 model had the best segmentation and detection effect in three scenarios,and Mask R-CNN-R101 had the worst effect.[Conclusion]Cascade Mask RCNN-R50 model can produce more accurately and effectively segmentation and detection of grape clusters in different planting scenarios,which can provide model support for automatic grape picking.