Study On Surface Defect Detection Algorithms Of Jujube Based On Machine Vision

Xinjiang jujube has extremely high nutritional and medicinal value,and is also known as “golden birthday dates”.They are favored by consumers and the demand for society is increasing,which has stimulated the enthusiasm of entrepreneurs to industrially manage jujube trees.At the same time,defects such as mildew,insects and cracks seriously affect the quality and value of the jujube and they must be sorted.Machine vision technology has the advantages of high efficiency,high precision,rich detection information,non-contact,etc.It has been widely used in the field of agricultural product defect detection and quality classification.However,the detection of defects on the surface of jujube is still being continuously explored and improved,and its recognition technology is more difficult,which is a difficult problem in detection and recognition.Based on the theoretical research and practical application of the predecessors,this study uses the experimental collection platform to realize the two kinds of data collection of static images and video of red dates.For the static red jujube image,the feature extraction based on the best combination of color and texture is studied.Finally,the suitable machine learning algorithm is used to detect the jujube with defects on its surface.At the same time,the convolution neural network can learn the differences between things autonomously,transform the original image data into a higher level and more abstract expression through network training,strengthen the characteristics of strong representation ability,weaken the advantages of irrelevant factors,and combine reinforcement learning.The method improves the detection and recognition rate of defective red dates.First,study the red jujube segmentation method and image preprocessing based on image processing technology.The machine vision image acquisition system is used to collect video and image data.The images captured by the camera,due to illumination problems,the acquisition system's own problems and environmental influences,will cause factors that hinder image segmentation,extraction and recognition.Therefore,the research work mainly uses geometric transformation,median filtering,image enhancement,and morphological operation for image preprocessing,which helps rotate,denoise and enhance the captured fruit image to prepare for subsequent image segmentation and feature extraction.Study the detection method of surface defects of jujube based on feature selection.Based on prior knowledge,determine the specific symptoms of the jujube defect.The ability of color and texture features to characterize defective red dates was studied,and the parameters with strong representation were determined for feature expression.According to the obtained feature vector,the appropriate classifier is selected and the defect detection of jujube is realized through sample training.The brightness(V)component of the HSV color space is selected as the illumination-reflection model input,and the surface brightness uniformity correction is performed to extract the color features of the R,G,B,H,S,V and brightness corrected images.At the same time,the local Binary Patter(LBP)and the Gray-Level Co-occurrence Matrix(GLCM)statistic are used as texture features.Defect detection is implemented by using a support vector machine classifier.Finally,the research on surface defect detection algorithm of jujube is studied based on convolutional neural network.The research proposes a target location method based on shortest path search between frames and an Ensemble-Convolution Neural Network(E-CNN).By establishing image coordinate system and image preprocessing operation,the position coordinates of each red jujube target in the image are obtained and mapped into the space coordinate system.Combined with the shortest path determination rule between frames,the target position coordinates are updated with the video time series.Pass and use this method to build data sets quickly and efficiently.Based on the “Bagging” integrated learning method,E-CNN builds a basic convolution neural network tree model through the training set,and then outputs the result according to each basic tree model,and obtains the final result of the model by “voting”.The experimental results show that the target location method using the shortest path search between frames has a positioning accuracy of 100%.At the same time,using the E-CNN model,the recognition accuracy and recall rate of the model reached 98.48% and 98.39%,respectively,and the classification accuracy was greater than the color feature classification model(86.62%),the texture feature classification model(86.40%)and the basic convolution neural network Model(95.82%).