Study On Determination External Defects And Classification Of Walnuts Using Machine Vision

Walnuts, which is one of fourth famous nuts in the world, has a lots of nutritional and economic value, and has a large of planting area and regions. In the process of production and processing, how to identify defective walnuts accurately from good ones classify the good walnuts, the recognition time and efficiency meet the requirement of on-line detection at the same time, are the important factors of influencing walnuts product automatic processing, food safety, and market competitiveness.This research is based on machine vision technology, the objective are achieved to recognize walnuts external defects and classify the good walnuts. The main contents of this research are choose the suitable critical components, according to photometric characteristics of external defects and characters of walnuts, which are include camera, light source, camera lens, image capture card, etc, design and construct acquisition and analysis system of RGB image base on machine vision system, ascertain the acquisition parameters of system of object distance, depth of field, spatial resolution, visual angle, background shooting, etc, base on orthogonal test. According to the feature of gentle transition of gray level distribution in the image for external defects and object of walnuts, design a adaptive dual threshold segmentation method which is suit to process on-line, extracted 20 demention initial characteristics for a defective region include geometry, texture, space, transformed into 9 demention optimized features, built 15 models based on bayesian, support vector machine, BP neural network take the optimized features as input, compared the recognition rate, unit discriminant time, stability, and training time of models to optimize the defect recognition modes. On the basis of discrimination results of defect recognition models, extracted 33 demention features of 4 types of geometry, texture, color, space, from the object region, used PCA method to extract principal components and obtained the principal components’ scores, with different number of sequential principal components for the input, built classification models of support vector machine and BP neural network, analyse the indexes of classification accuracy, unit processing time, stability, training time, etc,determine the number of input principal components and optimal model.The purpose of this research is explore the feasibility and adaptability of detecting walnuts external defects base on machine vision and processing on-line, provide the basis for designing and researching equipment of walnuts product. The conclusion of this research were that designed and structured machine vision RGB image acquisition and analysis system of walnuts based on the key apparatus 3-CCD area-array camera and ring diffuse light sources, white as the background shooting, used orthogonal test to optimize acquisition parameters, the system could aquire images with high signal to noise ratio, distinguishability, and 0.265mm/pixel for spatial resolution, basic meet the research requirement, especially shown up obvious effect in color revivification degree and compensation for characteristic region, it provide the thinking for developing of detecting equipment base on machine vision.The results of image segmentation for defective and object region shown that the proposed adaptive dual threshold method could accurately segment out regions of crack, damage, black spot and object, and the average process time was 0.5644s for a image under the environment of indows 7, Matlab R2010b, Intel i3-3110M CPU 2 X 2.4GHz.The results of building defects recognition model revealed that the 9 demention features could effectively describe defects, the RBF-SVM model have better comprehensive performance as the 9 demention features, the identification precision were 93.06% for crack,88.31% for damage,89.27% for black spot, and the total rate of correct discrimination was 90.21%, the unit identification time was 10’4s class under the environment of Windows 7, Matlab R2010b, Intel i3-3110M CPU 2×2.4GHz.The results of classification for good walnuts shown that, in the 33 demention characteristic space, the first 14 demention principal components cumulative score more than 90% as input built classification models of SVM and BPNN, among the models, the optimized model of SVM has the best classification result, the classification accuracy was 95.9%, unit process time was 10-4s class under the environment of Windows 7, Matlab R2010b, Intel i3-3110M CPU 2×2.4GHz, according to the results of classification for SVM model, the practical rate of correct discrimination was 84.58% compared with artificial classification result.In conclusion, the method, based on machine and support vector machine, could accurately identify defective walnuts and classify good walnuts, and could provide technical support for developing nondestructive equipment of other nuts or crops.