Recognition Method For Kiwifruit In Field

China is one of the largest countries in Kiwifruit producing all over the world.Nevertheless, the hand-picking work is a time-consuming, laborious work. With theaccelerating of labor force transformation from countryside to urban areas, Kiwifruitharvesting robots can help liberate the workforce, and improve picking efficiency largely,which will become a kind of inevitable trend.The primary task of kiwifruit picking robots is to accurately identify the kiwifruit in thenatural environment. Therefore it is one of the key techniques of picking robots to recognizeaccurately kiwifruits. This paper studied kiwifruit recognition based on the kiwifruits in thecomplex natural environment. What’s more, the paper realized the accurate segmentation ofthe target fruit area and the complex background area, and restored and recovered accuratelythe single fruits that were overlaid by others, or adjoined other kiwifruits. The main contentsand conclusions are as follows:(1) The paper studied the relative theories about robot vision, and analysed thecultivation forms of field kiwifruits and the various factors that influence the picking result.Moreover, the paper discussed the hardware and software structure of robot vision. Itexplained the composition of hardware, and the method of software implementation andimage gather.(2) This paper applied a single color channel, K-means clustering and chromaticaberration for kiwifruit segmentation. K-means clustering method selected a*and b*component of La*b*space. The chromatic aberration chose R-G channel. The results showedthat a single color channel could not be suitable for kiwifruit segmentation; R-G color methodcould not be a better segmentation for soil and kiwifruit, and twigs, stalk splitted poorly;K-means method extracted kiwifruit boundary information incompletely, and appearedserious over-segmentation. Therefore, the application of these traditional fruit segmentationmethods were not very satisfactory.(3) The paper present Adaboost algorithm for fruit segemention. It built six differentweak classifiers using separately one or more color channels in three color spaces which wereRGB、HSI、La*b*. The strong classifier was generated with300training samples. Then thepaper verified the precision of weak and strong classifiers with655test samples. The result showed the precision of the strong classifier was higher than any weak classifier, and theresult was94.20%. This algorithm can eliminate most of the backgrounds, and thesegmentation result had improved greatly than color method and K-means clustering method,which reached an ideal recognization effect.(4) Compared and analysed different algorithms fruit recognition rate. Choosing80images for Adaboost, R-G color method and K-means to image segmentation. Statisticingrecognition rate after removing different types of residual noise by range of morphologicalmethods. Compared the results of the three methods, the results showed that Adaboostmethod identified the maximum number of kiwifruit, the recognition rate up to96.7%.(5) The experiment restored the kiwifruits in images. Through comparative analysis,selecting Canndy to extract edge contour of Kiwifruit. Calculating the curvature of the fruitcontourto obtain the smooth contour curves, then using the least squares method to achievethe fruits of recovery.