Research On Vegetable Classification Based On Support Vector Machine

The quality of food circulation is related to the national economy and people's livelihood,and it is an important part of China's national economy.However,the sorting of vegetable circulation which is an important part of food circulation still has many problems such as artificial,low efficiency and low degree of automation.The application of automatic sorting method based on machine vision to vegetable logistics sorting can solve the above problems very well,among which the most key technology is vegetable recognition algorithm based on machine vision.The main work of this paper is to study the recognition algorithm of automatic vegetable sorting on the basis of self building vegetable image set by collecting images,and carry out in-depth research from three aspects: image preprocessing,feature extraction and classification recognition.(1)Image collection.Two image sets,named image set 1 and image set 2,were constructed based on the vegetable pictures collected by a micro-single camera.Image set 1 is the largest proportion of vegetables consumption in China,including cucumber,potato,tomato,white radish and cabbage.Another kind of image set divides potatoes into four types based on their quality: sprouted,green,damaged and intact.(2)Aiming the condition that pictures collected have obvious contrast between foreground and background such as dark background and obvious scratches,a kind of color image segmentation method based on HSV is puts forward whose main method is color channel separation,threshold segmentation and image addition,etc after image size changes and smooth filtering.(3)Aiming at the problems of illumination invariance and poor noise robustness of texture features extracted by the complete local binary pattern,an adaptation threshold complete local binary pattern with direction is proposed.According to the discriminant conditions,one of the adaptive neighborhood pixel median and center pixel values is selected as the threshold of LBP coding.Then,the amplitude of the difference between the neighborhood pixel and the threshold is calculated,and the binary mode of the local neighborhood is calculated from the smallest point to the largest point.By running the algorithm on the common texture database,the probability of two kinds of threshold is calculated,and the necessity of adaptive threshold is proved.The experimental results show that the algorithm has invariance of rotation and illumination,and is robust to salt and pepper noise,which can describe the texture better than similar algorithms.(4)An improved weighted feature fusion method based on classification accuracy is proposed,which regards joint feature vectors obtained by combining dat-clbp and color histogram with accuracy.as feature descriptors,and considers a support vector machine based on radial basis kernel function with parameter c=10 and ?=5 as the classfier.This method was verified experimentally on the self-built image set,and compared with single feature vector,serial fusion feature and equal weight fusion feature method,which proved the method had higher stability and recognition rate.