Research Of Plant Leaf Recognition Algorithms Based On Centrist And Implementation On Mobile Platform

Plant is closely related with human among all organisms living on earth, has plenty valuein foodstuff, medicine and industry. It also plays an important role in environmentalprotection. As a basic work of botany research and agricultural production, the classificationand identification of plant species has vital significance for protecting plant, distinguishingplant species and exploring the genetic relationship between plant species. The traditionalmethods of recognition plants mainly rely on manual, which needs plenty of domain expertise,costs heavy workload, is inefficient and difficult to guarantee the objectivity and accuracy ofthe classification. Along with the rapid development of information technology, theintroduction of computer vision, pattern recognition, database and other technology came into help people recognize the specie of plant which is very in need. The images of plant leafare much easier to acquire relative to other parts of the plant. The leaf color, shape and veinall can be used for classification.Based on summing up the domestic and oversea relative researches, this paper chose plantleaf image as research object, proposed a series of novel techniques for plant leaf recognitionby using CENTRIST(CENsus TRansform hISTogram) visual descriptor. Spatial principalcomponent analysis of census transform histograms and improved spatial census transformhistograms algorithms are used to extract features, which are used by some classifiers toidentify plant leaves. At last we implement a mobile application running on Android devices,achieve the fast recognition of plant species from leaf images.The main contributions are as follows:(1) Acquisition and preprocessing of leaf image datasetThis work uses Flavia plant leaf database as the experimental dataset. The databasecontains32categories plant leaves with55-77samples for each category. All of the plant leafimages should be preprocessed before being used, such as size transformation, unifyingbackground color and adjusting the location of the leaf images.(2) Researches on feature extraction algorithms of plant leavesThis study does research on fundamental of CENTRIST and some feature extractionalgorithms based on CENTRIST, such as CENTRIST, sPACT(spatial Principal component Analysis of Census Transform histograms), and so on. These methods are applied to extractfeature of plant leaves in order to get feature vectors, then two kinds of classifiers areemployed to obtain recognition results, at last the experimental results will be analyzed.(3) Improvement of CSNTRIST algorithmThis study improved and optimized CENTRIST algorithm based on the study ofCENTRIST and sPACT, then derived IsCT(Improved spatial Census Transform histogram)algorithm.. From experimental point of view, this study analyzes the correct recognition ratesusing different algorithms, and determines the most suitable system parameters for mobiledevices.(4) Recognition of plant leavesThe experimental dataset consists of32plant leaf classes with a total of1907colorimages. Four algorithms are used to extract features of the leaf images, and then these featuresare applied to two kinds of classifiers to obtain recognition rates. Experimental resultsdemonstrate that the average correct recognition rate reaches97%, and meets the designrequirements, so CENTRIST can be used to identify plant leaves. The system facilitates leafrecognition with a good performance. The most important thing, it is very easy to implement.