Research On The Shape Analysis Based On Fourier Transformation And Application In Leaf Images

Plants are closely related to Human’s daily life.Based on image analysis, researching how to classify and retrieve plant leaves has practical significance meaning of protecting the plant resources, mining the plant evolution rules and exploring the genetic relationship between plants.The traditional method of identifying plants mainly depends on artificial record, which has problems of low efficiency, time consuming, high requirement of experiential techniques, and large deviation. As a hot technology in the field of intelligent information,content-based image retrieval technology (CBIR) has provided a feasible solution for plant classification and retrieval.This paper focuses on two core issues:plant leaf image feature extraction and feature matching.The author systematic analyses various techniques based on the content image retrieval,then summarizes the advantages and disadvantages of various shape description methods for image retrieval.The main research contents are as follows:(1).We research the methods which using Fourier coefficients to describe the shape characteristics, and detailed analysis various of Fourier descriptors to achieve normalization process of scaling invariance, rotational invariance and translational invariance.According to the normalization criteria for scale,we attempt to standardize in the space scale, and put forward a improved complex function of the coordinates Fourier descriptors (GCCFD),a improved center Fourier descriptors (GCDFD) and a multi-scale Fourier descriptors (DFD).Furthermore, we derive the generation of new Fourier descriptors in theory.(2).The test dataset adopts 100 different kinds of plants leaves which consists of 1000 images, and which is collected form the condition of nature world.Then we design retrieval scheme using the methods of CCFD method, GCCFD method, CDFD method, GCDFD method and DFD method.The achieve the target of separating the object and the background, we use a series of preprocessing process, which include of decoloration, object contour extraction.(3).The thesis employs the block distance as the similarity measure standard to realize retrieval function.We not only evaluate the system with conventional PVR curve performance criteria, but also propose bull’s-eye measure to assess image retrieval performance in multiaspect.The experiments show that the GCCFD method, GCDFD method and DFD method have some advantages over traditional methods.Using these methods, the retrieval performance has improved in some extent. Moreover, the experiments verify that using Fourier descriptors to retrieve has advantages of low time complexity, rapid convergence performances and robustness.