Research On Automatic Graded Technology Of Spherical Fruit And Vegetable Based On Machine Vision

China is a large agricultural country, vegetable industry is the second industry and fruit industry is the third industry in the planting industry.Using machine vision technology,vegetables and fruits’ quality were automatic evaluated, it can improve the objectivity of fruits and vegetables’ grading,reduce labor intensity of employees, and it also can improve the commercial processing of fruits and vegetables, promote the international competitiveness of fruits and vegetables.The work mainly around spherical fruit and vegetable’s automatic grading technology, and it works roughly in the following order: spherical fruit and vegetable’s images acquisition, image preprocessing, feature extraction, feature vector’s optimization, classifier to classify, after these processing,spherical fruit and vegetable grading achieve the desired goal. The main work and contributions are described as follows:(1) Spherical fruit and vegetable image database is built,a image significant region extraction method is proposed. In order to simulate machine visual observe images, a spherical fruits and vegetables image acquisition box has been designed, some kinds of multi-stage spherical fruit and vegetables images are collected,including apples, persimmons, pears, tomatoes and oranges, each class has three levels. Because the collected images’ target size are too big, the background are too much, the centroid of target is not centered, so a spherical fruit and vegetable image preprocessing scheme is proposed, the scheme mainly contain size magnification, extract target’ s contour and find the centroid, intercept the target area as much as possible. Finally, spherical fruits and vegetables’ image databases have been built.(2) For the classification problem of extracting spherical fruits and vegetables’ texture feature, using blocked local binary pattern as the spherical fruits and vegetables’ texture feature for classification is proposed. In order to reduce the local binary pattern’s characteristic dimension, taking full account of dimensionality reduction and spatial relationships, we let block mind to original local binary pattern,then the blockd local binary pattern is obtained, which can avoid confusing spatial position, at the same time the dimension of feature vector is reduced.Under the same experimental conditions, using blocked local binary pattern for spherical fruits and vegetables’ grading is more effectively than original local binary pattern.(3) For the problem of generating histogram of oriented gradient is length, a improved histogram of oriented gradient shape feature extraction algorithm is proposed. On the basis of the original histogram of oriented gradient simplifying the details of the dividing unit, improved histogram of oriented gradient algorithm is proposed, which can keep the histogram of oriented gradient’s advantages,such as can describe local shape,can restrain the translation and rotation effect,what’s more characteristic dimension is greatly reduced.Under the same experimental conditions, improved histogram of oriented gradient’s grade accuracy is higher than the original histogram of oriented gradient.(4) To solve the problem of single feature describe image content is not comprehensive, A combined characteristics of spherical fruit and vegetable grade method is proposed.For the weights assignment problem of combined characteristics, support vector machine’s independent distribution feature weights and adaptive feature weight distribution is studied, feature weights’ optimization of multi-layer perceptron is also studied.In support vector machine, we focus on analysing and discussing the generalized multiple kernel learning. Experiment results show that the adaptive allocation features weight distribution method generalized multiple kernel learning is more intelligent and efficient than independent distribution feature weights. In multilayer perceptron, selecting biogeography based optimization method to optimize feature weights, the multilayer perceptron input datas are the CMs, QSGLCM, IHOG’s dimensionality reduction combination features. By comparing biogeography based optimization method to particle swarm optimization method for optimizing feature weights, the results shows that biogeography based optimization method can obtain better feature weights, and can improve the grading accuracy of spherical fruits and vegetables effectively.In this dissertation, through horizontal and vertical comparative experiments, the best automatic graded method of spherical fruit and vegetable based on machine vision is got,the grade accuracy achieves above 90%.