Research On Apple Grading Detection Algorithm Based On Machine Vision

The output value of fruit can be improved to some extent through post-natal treatment and post-production processing,and the post-natal promotion of the value of fruit is called the commercial treatment of fruit,and the post-natal grading stage of fruit is an important part in the commercialization of fruit.However,the fruit grading technology in our country is relatively backward at present.The main grading methods for domestic fruits are manual grading and mechanical grading,there are many deficiencies in these two classification methods: grading standard instability,unideal adaptability,easy to damage fruit,low classification efficiency,and so on.In recent years,China has begun to study the automatic classification of fruits based on machine vision technology,the method mainly collects the target image through CCD cameras,then through computer analysis and other techniques to analyze the fruit image and come to a comprehensive conclusion.This method can be used to detect and classify external indicators of fruits in an objective,real-time and non-destructive manner.The apple with the first yield in China was selected as the grading target and studied the apple grading detection algorithm based on machine vision technology in this thesis,In the algorithm,some image detail optimization methods were added to make the result more stable and ideal.At the same time,it had certain application research value and the algorithm was verified and analyzed in the Matlab platform.Moreover,the algorithm was not only suitable for detecting a single target in an image in this thesis,but also can perform recognition processing when there were multiple targets in an image,it would provide early algorithm support for apple sorting robot or picking robot in the future.The main research content of this thesis are as follows:(1)Selecting the background color suitable for the apple,and completing a series of preprocessing such as filtering and denoising,grayscale contrast enhancement.In the stage of image denoising,a combination of adaptive median filter and wavelet transform was chosen to perform the denoising process,and the PSNR standard was used to evaluate the denoising quality of commonly used filtering methods.In orderto increase the contrast between the apple and the background to a greater extent,a contrast stretching function method was applied to enhance transformation through analyzing several transform methods in the image enhancement process.(2)In the stage of image segmentation,it was analyzed and compared between one-dimensional OTSU threshold method and two-dimensional OTSU method,considering the algorithm running time and segmentation effect,one-dimension OTSU threshold method was selected for segmentation of apple and background.Then,aiming at the problem of influencing the effect of image segmentation due to uneven lighting background and combining the advantages of open operations,The top-hat transformation detail algorithm was introduced to optimize the segmentation method to make it more generalized,and the improved Canny operator was used to identify continuous and smooth apple contours.(3)Apple's feature extraction in size,shape and color,and statistics of related data.The advantages and disadvantages of various methods were taken into consideration in the extraction of apple size features.Finally,using the largest inscribed circle method to calculate the diameter of apple,the classification of size was more convenient.In terms of apple shape features,using the roundness value to describe and grade the shape characteristics of apples.In terms of apple texture features,surface textures were measured using descriptors such as second order moment,third order moment,and entropy based on center order moment;In apple colors,The RGB color space model of apple image was converted to HSI model,and the color pixels of the H component of the apple sample image was identified,and the color grade of the apple's surface was divided by calculating the apple red coloring rate.Experiments had proved that these feature extraction methods were highly feasible,the accuracy rate was above 90%,and there was little difference with the results of artificial classification,which meets the classification requirements.