Research On Vision System Based On Support Vector Machine Using In Apple Harvesting Robot

Researches of critical technologies on fruit picking robot,not only have a certain utility value to adapt to market demand,to reduce labor intensity and to improve economic benefit;but also have important practical significance for tracking new technologies in agriculture of the world,promoting the process of agricultural modernization in China.This research is supported by the National High Technology Research and Development Program of China(863 Program,No.2006AA10Z254),we take mature apples in natural environment as study objects,study from the actual situation of the apple harvesting,using apple's color,shape and other characteristics,combining with the technology of Support Vector Machine to segment and recognize images,and the main work has been accomplished as followed:1.Design of Experiment System.The hardware structure and software platform of the fruit picking robot vision system has been built;after analyzing the environment of this study,the experiment precept of image acquisition has been designed using of VFW modules of Windows system,which makes up disadvantages such as long acquiring time and bad real-time performance and which takes apple images acquired in different light conditions as the main data of image analysis.2.Image Pre-processing.There exists noise disturbances in acquired images,Gaussian filtering method and median filtering method which is in common use,can make the edge vague,flood the characteristics,and bring about difficulties for analyzing.Color image vector median filter algorithm used in this study,not only can remove noise,highlight the apple fruit;but also can overcome the disadvantage such as bad performance in edge maintenance in traditional filtering algorithms,that is, have better performance in edge and detail maintenance;furthermore,in the process of filtering,the central pixel of the window is replaced by the distance which is the shortest,not replaced by the synthetical vector,so their properties will remain unchanged.This is one of the critical technologies in this study.3.Image Segmentation.In natural environment,there are four light conditions:front lighting, back lighting,fruit in the shade and cloudy.Single threshold method,which is used as segmentation algorithm commonly,isn't influenced by the homogeneity of light a lot,but can't segment the fruit entirely.In this study,the Hue-Histogram Statistic Double-threshold Algorithm Based on HIS Color-space and Region Growing Algorithm are put forward,which extracts the chromaticity component that does nothing with intensity in HIS model,rules out the influence for image quality resulted from different lightness,and the algorithm is simple,time consuming is short,segmentation regions are complete.This is another critical technology in this study.4.Feature Extraction and Fruit Recognition.Artificial neural network method which is in common use,requires a large amount of training sample sets,while the template matching method needs to establish a database and expert system,so it will result in a heavy workload.The Support Vector Machine constructs the determined optimal hyperplane under the principles of minimum frame hazard,which can train the learning machine in the case of limited samples,and which also overcome the disadvantage that needs a large amount of samples in neural network method.In this study,the model recognition method,which is based on Support Vector Machine combined color and shape properties,is used.Experimental results indicate that the performance of SVM is superior to neural network that in common use,the effect of recognition is better,the accuracy rate is higher,and it shows predominance in small samples;recognition rate of apple fruit based on RBF SVM of color and shape properties is higher than that of only using the color or shape properties.This is the third critical point in this study.5.Experimental verification.This study validates the reliability and practicality of the algorithms from accuracy,real-time and stability,takes apples,which are red not bagged,red bagged and yellow in ideal states,as the researching objects.Experimental results show that,apple recognition algorithm based on SVM not only has better recognition effect for red apples,but also reflect predominance compared with traditional recognition method based on color properties for yellow apples;for accuracy,this method can recognize and locate the mature fruit in images,the correct recognition rate reaches to 93.3%;for real-time,the recognition time is about 100ms,which can meet the requirement of the real-time of harvesting robot and the adaptive adjustment of following images; for stability,in the process of continuous picking,the coordinates of picking point changes a little, which ensures the moving path is shortest and the curve is smooth,reflecting a good robustness.