Research On Leaf Recognition Based On Structure Integral Transform

Leaf recognition is an important application of computer vision and is of great significance in modern agriculture and biological sciences.Due to the small inter-class differences and large intra-class variations of the leaf shapes,coupled with interference factors such as leaf self-overlap,leaf incompleteness and image noise,the issue of effectively recognizing the leaf shape is not still well-addressed.Compared with the traditional contour-based shape descriptor,the region-based Structural Integral Transform(SIT)can extract the structure relationship from all the pixels in the leaf region in the hierarchical description framework.These features are more robust to the self-overlap,contour defect of leaf shapes and image noise.However,the core line integral in SIT is only to calculate the length of the integral line,which cannot effectively describe the details of leaves,and it does not use the key grayscale information of leaf images,so its ability to characterize the leaf shape is not enough.In addition,SIT extracts features according to artificial rules,which are redundant to a certain extent.Dictionary learning can learn common shape patterns from different kinds of leaves,obtain sparse and discriminative features,and improve the accuracy of leaf recognition.Therefore,this thesis studies how to extract the structure relationship and detail information of the grayscale leaf shape simultaneously based on SIT,and how to use dictionary learning to reduce feature redundancy and construct a more discriminative leaf shape descriptor.The main research work of this thesis is as follows:(1)Aiming at the key structure and detail characteristics of grayscale leaves,the grayscale SIT based on multi-scale piecewise line integral(MGSIT)is proposed to construct a multi-scale structural description of grayscale leaf.Firstly,based on the whole leaf shape,a 2D dissecting structure is designed to construct a hierarchical structured description framework.Secondly,integrating the grayscale shape over the 2D dissecting structure.Next,dividing the integral line in the 2D dissecting structure by multi-scale bisection operator,and integrating the grayscale shape over the integral line segment.Higher scales mean more times of bisection of integral lines,shorter segments of integral lines,and more detailed features can be extracted from them.Finally,invariant descriptor is constructed to realize leaf classification.The experimental results show that the proposed method can effectively characterize the leaf,improve the performance of cultivar leaf recognition and species leaf recognition,and can be well generalized in general shape recognition tasks.(2)In this thesis,the dictionary learning method based on SIT(DLSIT)is proposed to construct sparse and more discriminative leaf shape descriptor for different classes of leaves with common shape patterns,which can reduce the feature redundancy of SIT to some extent.Firstly,dictionary learning method is used to learn dictionary on each scale from the MGSIT features of leaves in training set.Secondly,the sparse representation of the leaf on each scale is obtained by using the dictionary on each scale,and the sparse representation of each scale is counted to form the final description of the leaf.Finally,the leaf recognition is realized by 1NN classifier.The experimental results show that the proposed method not only outperforms MGSIT,but also reduces the feature redundancy.The main contributions of this thesis include proposing multi-scale piecewise line integral strategy and constructing leaf shape descriptors to extract structural and detailed information of grayscale leaves,and introducing dictionary learning to obtain sparse and more discriminative leaf representation.The experimental results show that the proposed method can improve the recognition rate of cultivar leaf and species leaf with rich internal details,and can also play a greater advantage in recognition of general shape with inaccurate contour information or large deformation.