Research On Leaf Classification By Using Multi-scale Texture And Contour Feature Fusion

The classification of cultivars can promote the establishment of correspondence database about plant phenotype and trait,and gradually realize to control the process of crop breeding by information technology,thus promoting the realization of molecular breeding.Most of the existing plant classification methods take leaves as the study object to classify different species of plants,and have achieved remarkable research results,but there are few studies on fine-grained leaf classification of different cultivars of the same species.Fine-grained leaf classification is a finer secondary classification based on species classification.The high similarity in leaves brings great challenges to classification of cultivars.Therefore,how to describe the detailed information contained in leaf images more finely and extract distinctive features is the key issue of fine-grained leaf classification.To address the above issue,a multi-scale texture and contour feature fusion classification method is proposed in this thesis,which taking leaves of soybean as the study object,aiming at exploring the fine-grained leaf classification and realizing plant cultivars classification.The main work is as follows:(1)The soybean leaves datasets were established to provide data support for the study of fine-grained leaf classification.Soybean is the main cash crop in China.It is widely planted and the leaves are easy to collect.Three different soybean leaves datasets were established,they are Multi-Parts Leaves(MPL),Single-Part Leaves(SPL)and Multi-Growth-Stage Leaves(MGSL).MPL and SPL datasets are used in this thesis,and MGSL dataset is collected for other study of our research group.(2)Leaves of different cultivars have very high similarity in shape.Thus,the features need to enhance the ability to express local details of leaves.Therefore,a multi-scale line-based local binary pattern(MLLBP)texture descriptor is proposed to extract the detailed information contained in leaves.Firstly,the contour is cut by straight line connecting two contour points,and different scales are defined based on the contour curve segments with different lengths after cutting which form a multi-scale frame.The scales with longer curve segment extract the global features of leaves,express the overall information of leaves,the scales with shorter curve segment extract the local features of leaves,and express the details of leaves.Then,the line-based Local Binary Pattern(LLBP)is proposed to improve the computational complexity of LBP and get the robustness to background noise,which get the pixels on straight lines and encoded them with LBP.Combining multi-scale framework and LLBP,named as MLLBP features,can describe leaves from global to local.At the same time,according to the spatial distribution characteristics of contour points,multi-scale contour feature is extracted to describe shape information.Finally,the invariance of scaling,translation,rotation and of features are analyzed and proved in detail.(3)To achieve the fusion of texture feature and contour feature,feature weight learning algorithm is used to adjust the weights of the two features for achieving feature fusion,and SVM is used to as the classifier.Firstly,SVM classifier is constructed for texture feature and contour feature respectively.Then,the weight of each feature is automatically learned according to the training set.Finally,the classifier based on the two features is combined into a complex classifier by using the feature weights.The experimental results show that this method has higher classification accuracy compared with texture feature,contour feature and directly concatenate feature.