Research On Grading Method Of Red Fuji Apple Based On Machine Vision

The apple industry is a very important part of the agricultural and forestry economy,although the production and consumption capacity of apple in China is in the forefront of the world,China’s post-harvest processing technology is backward.Especially the commercialization grading technology of apple is still insufficient compared with the advanced technology of foreign countries.At present,the level of automation in apple grading is still relatively low,the grading effect is not ideal,and apple products are lack competitiveness in the international market.In this paper,the external quality grading method of Red Fuji apple is studied based on machine vision technology.The main research contents are as follows:(1)This paper analyzes the RGB and HSI color models of apple images firstly.And a background segmentation method of "preliminary segmentation combined with Otsu threshold segmentation" is adopted,which can effectively reduce the amount of calculation and realize image background segmentation quickly and accurately.In the process of gray image smoothing,comparing the denoising effects of the neighborhood average with the median filter.And the experiment shows that the median filter can not only effectively remove noise,but also better protect the information of the image on the edge.Based on the principle of morphological filtering,by deleting the small area noise,hole filling,and morphological operation,the noise generated in the process of image segmentation is removed,and a binary image of an apple with a complete region and smooth edge is obtained to complete the sample image background segmentation.(2)The external quality features of the apple are extracted,such as color,shape,and diameter.When extracting the color features of apples,a large number of HSI color models of apples of different grades are analyzed.And then this paper takes the hue distribution of images as the color features of apples.Besides,the red part of the apple is segmented,and the ratio of the red part to the whole surface of the apple is calculated as the color feature of the apple.This paper uses an 8 neighborhood nonlinear filtering algorithm to find the edge pixels in the apple binary image and then the apple edge contour is obtained.The roundness and the ratio of transverse diameter to longitudinal diameter are calculated to describe the shape characteristic of the apple.It founds that the latter is more in line with the actual situation of apple.For the characteristics of the apple fruit diameter,this paper calculates the length of the smallest circumscribed rectangle and the diameter of the smallest circumscribed circle,and compared them with the manual measurement.The result shows that the minimum circumscribed circle method is similar to manual measurement.(3)The support vector machine based on genetic algorithm and the convolution neural network are established in this paper.On the support vector machine,9-dimensional vectors are extracted from each apple image as the input of the support vector machine.Then,a support vector machine for four classifications is constructed by the "one-to-one" strategy.By comparing the classification results of SVM with different kernel functions,the Gaussian kernel is used to map the nonlinear data.In the training process of SVM,the genetic algorithm is used to find the global optimal parameters,and the test result shows that the grading accuracy of the support vector machine optimized by the genetic algorithm is 94.64%.On the convolution neural network,a 10-layer convolution neural network is constructed,and the apple images are preprocessed and enhanced as the input of the convolutional neural network.Then the training parameters are continuously optimized through a lot of experiments,and the network model with the highest accuracy is saved.The test result shows that the grading accuracy of the convolutional neural network model is 93.28%.Finally,another 240 apple images are obtained to test the two models.The result shows that the grading accuracy of the support vector machine optimized by the genetic algorithm is higher than the convolution neural network.(4)This paper studies the apple grading system which mainly includes the system communication and the design of the software.It can carry out apple’s real-time acquisition,grading and statistics,and output relevant information to the software interface.