Study Of Feature Extraction Based Genetic Characteristics And Species Identifiction Of Weed Seeds

Weed seeds have a great harm to agricultural production. Every year, it causedhuge losses in crop production all over the world. Strengthening the weed seedquarantine and identification research is very important to automatic identification ofweed seeds. Because of the wide variety of weed seeds, as well as its morphologicalchanges greatly in different growth period, it means the identification aims to solve akind of high dimension, massive target recognition and classification problems.In this paper, mature leguminous weed seeds are studied as the research object,and the weed seed morphological characteristics and HU moment invariant feature isextract based on the biological stability of genetic characteristics. A weed seedclassification and identification system is designed, through the analysis of therelations between these features. The experimental results show that the system canimprove the recognition rate of weed seeds. The main research contents are asfollows:Traditionally, magnifier or microscope is usually used for visual classification,where higher labor intensity and technical expertise is required. In this study, moderninformation processing is used to realize automatic recognition and classification ofweed seeds, so as to improve the recognition rate and to reduce the manual operationtime.Weed seed image pre-processing is used, including the image histogram analysis,image segmentation, and image denoising and edge detection. Through thecomparison of various algorithms, an edge extraction method is proposed which isbased on morphological processing to extract interested contour information of theweed seeds, so that the weed seed can be separated from background.Leguminous weed seeds’ main parameters can be extracted based on biologicalstable genetic characteristics. These parameters mainly include9MorphologicalCharacters and7HU moment invariant feature,16features have the characteristics of translation, rotation and scale invariance. Due to the wide variety of weed seeds, itinvolves a problem of vast category and high-dimension data classification andidentification. So it need reasonable dimension reduction for the high dimensional andmassive data. This paper mainly uses the method of principal component analysis,using4principal component feature values to replace9feature values, combiningclustering analysis methods, to classify these high dimensions, large data sets.This paper evaluates the relation between the image features and recognition rateby BP neural network. The experimental results show that, the16characteristics, dueto high-dimension, have the highest recognition rate. The recognition rates of PCAdescending method differs little from16characteristic value recognition rate. As theBP neural network is easy to fall into local minimum value, the overall identificationrate is not high. Select the RBF kernel function for SVM modeling, the results ofidentification show that the identification rate increased significantly. To the weedseeds large-scale training sets, SVM’s optimizing speed, classification speed and theselection of parameters are not perfect. When using the method supporting vectorregression to training these samples, the SVR regression model has a betterperformance than the former two.