Research Of Deep-Learning-Based Branch Segmentation And Label Noise Suppression On Pepper Images

The development of China’s economic enters in a new normal,and the development of agriculture and rural areas facing with unknow challenges and opportunities.As the economic crop in the southwest of China,achieving automatic harvesting of Sichuan pepper is important.With the development of automatic technologies,most of the crops have been realized automatic harvesting at the oversea.The combination of image process method,communication technologies,and the robotic technologies is the future direction.Aiming at the problem that automatic harvesting of pepper in orchard,this paper conducts an in-depth study on the deep-learning-based method for foreground branch segmentation in orchard.The main contents are as following:(1)we summarize the domestic and international status of the image segmentation in agricultural field,conclude the difficulties of the branch segmentation for pepper in the outdoor sense,study the basic theory of deep learning related to the image segmentation,and analyze the key elements of the branch segmentation and label noise,and provide an feasibility solution.(2)Aiming at the problem that the previous model is difficult to segment the foreground branches in natural scenes,we propose a multi-modal and channel attention segmentation model(Spacial and channel attention,Attention＿SC)for RGBD images,which can effectively extract the representation of the foreground branches to achieve an accuracy segmentation.Firstly,the dual encoders are adopted to extract the features of the RGBD images,which overcomes the limitations of previous methods that only use a single feature;secondly,a multimodal spatial attention mechanism is constructed to fuse these two features;Then,the channel attention mechanism is adopted to extract branch features.Finally,the improved model is named as Attention＿SC.Theoretical analysis and experiments show that the segmentation accuracy of the proposed model is 70%.Compared with the existing segmentation network,the accuracy and the intersection and union ratio are increased by 14.72% and 17.65%,respectively.The proposed model can segment satisfactory foreground branches.The spatial weight both considers the appearance and distance information,which can suppress the interference of background branches,pay more attention to the shape information of the branches,and avoid the interference of detailed outlines of branches.(3)Aiming at the label noise generated in the annotation of the pepper branches,a label noise suppression algorithm based on loss function(Aleatoric uncertainties loss,Auloss)is proposed.Firstly,the weight of the branch segmentation model is changed to distribution,and the mean and sample variance of the sample are output at the end of decoder as the branch segmentation result and the sample uncertainty;then a new loss function is constructed according to the sample uncertainty,to Suppressing the negative impact of label noise in RGBD pepper images on network learning.Theoretical analysis and experimental results show that,for the Sichuan pepper dataset,the intersection ratio of the branch segmentation based on the loss function in this paper is 80.19%,which outperforms the loss functions adapted in other methods.The segmentation robustness of the network is improved,and the model in this thesis can provide the visualization description of the aleatoric uncertainty and model uncertainty.