Tree skeleton extraction plays a fundamental role in reconstructing both biological and structural models of trees. However, traditional approaches can be ineffective and problematic in guaranteeing the topological correctness and centeredness of the tree skeleton when the tree point clouds contain noise and occlusions. To overcome this limitation, we present a tree skeletonization method to generate topologically correct and well-centered tree skeletons. We extract an initial skeleton from the tree point clouds via an octree and level set method, use cylindrical prior constraint (CPC) optimization and the estimated radii of branches to yield corrected positions of improper joints, and finally obtain updated skeletons with improved smoothness. The good centeredness of our proposed method is intrinsically achieved by (1) exploiting the cylindrical shape prior and calculating the CPC in the local neighborhood and (2) feeding the prior knowledge regarding the radii of tree branches into a topology refinement algorithm to yield near-optimal estimates of the positions of the skeleton points. To evaluate our method, we construct a novel tree point cloud data set with known ground truth and propose three quantitative assessment methods: skeleton points deviation (SPD), bifurcation points coverage (BPC) and endpoints coverage (EPC). The quantitative assessment and visual assessment show that the proposed method clearly outperforms traditional ones in terms of topology correctness and centeredness of the extracted tree skeleton.

Deep learning has recently shown promising results in plant lesion recognition. However, a deep learning network requires a large amount of data for training, but because some plant lesion data is difficult to obtain and very similar in structure, we must generate complete plant lesion leaf images to augment the dataset. To solve this problem, this paper proposes a method to generate complete and scarce plant lesion leaf images to improve the recognition accuracy of the classification network. The advantages of our study include: (i) proposing a binary generator network to solve the problem of how a generative adversarial network (GAN) generates a lesion image with a specific shape and (ii) using the edge-smoothing and image pyramid algorithm to solve the problem that occurs when synthesizing a complete lesion leaf image where the synthetic edge pixels are different and the network output size is fixed but the real lesion size is random. Compared with the recognition accuracy of human experts and AlexNet, it was shown that our method can effectively expand the plant lesion dataset and improve the recognition accuracy of a classification network.

This paper proposes a deep learning model for the classification of citrus canker that overcomes the shortcomings of traditional approaches, and the scarce number of available images for training, which have been subject to the overfitting limitation. To address the issues, we propose two approaches, namely, the feature magnification and objective breakdown optimization, to augment datasets, and emphasizes meaningful features and prevent the model from overfitting noise signals. The first approach mimics the distribution of positive samples with a generative model based on deep convolutional generative adversarial networks. The second approach updates different parts of the model with optimization objectives, whereby back-propagated error signals are no longer the only signal for updating parameters. To validate the proposed approaches, we present theoretical proofs to justify the correctness of our methods and conduct extensive case studies and experiments to show that the proposed approaches clearly outperform traditional approaches on the classification of accuracy and efficiency of small training sets. In this paper, a methodology is proposed to generate a general model. The model can be applied to other bio-medical applications, where the scarcity of visual samples makes it difficult for a normal deep learning model to work without overfitting.

Jinpeng Zhang, Baijun Ji, Nini Xiao, Xiangyu Duan, Min Zhang, Yangbin Shi, Weihua Luo. Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021. 2021.

Clustering analysis is an unsupervised learning method, which has applications across many fields such as pattern recognition, machine learning, information security, and image segmentation. The density-based method, as one of the various clustering algorithms, has achieved good performance. However, it works poor in dealing with multidensity and complex-shaped datasets. Moreover, the result of this method depends heavily on the parameters we input. Thus, we propose a novel clustering algorithm (called the MST-DC) in this paper, which is based on the density core. Firstly, we employ the reverse nearest neighbors to extract core objects. Secondly, we use the minimum spanning tree algorithm to cluster the core objects. Finally, the remaining objects are assigned to the cluster to which their nearest core object belongs. The experimental results on several synthetic and real-world datasets show the superiority of the MST-DC to Kmeans, DBSCAN, DPC, DCore, SNNDPC, and LDP-MST.

Houquan Zhou, Yang Li, Zhenghua Li, Min Zhang. Findings of the Association for Computational Linguistics: ACL 2022. 2022.