Detecting Green Citrus Fruits On The Trees Based On Vision

The detection of fruits on the trees have attracted many researchers worked on image processing and agricultural engineering with the development of image processing and machine vision technology.It is very important to make full use of information and intelligent technology to monitor the growth of crops dynamically in fruits production and efficient management of the orchard,such as fertilization,irrigation and spray.And it is of great significance in crop-quality improvement,yield estimation and automatic harvest.The key point of early yield mapping and efficient management of the orchard to achieve automatic and intelligent fruits production is detection of green fruits on the trees.Detecting immature green fruit is still a more challenging task because of the lack of robustness of existing methods.Taking citrus as an example,the algorithm is divided into two parts: extraction of regions of interest(ROI)and hierarchical contours analysis(HCA).Firstly,the regions of interest in the image will be extracted by two methods separately.The first method takes the local binary pattern(LBP)features as the texture features for classification.The second method detects the maximum stable extremal regions(MSER)to extract the regions of interest in the image.Then,the hierarchical contour map will be set up and the hierarchical contours will be extracted by HCA based on annular illumination distribution of fruits which is first proposed in this paper.And multiple levels of contours will be extracted and fitted with the circular Hough transform(CHT).Finally,multiple fitted circles would be merged into one target which is considered as the last detected citrus fruits.The next paragraph is the summaries of the main contents and experimental conclusions.Contents:(1)Extraction of regions of interest based on LBP features.Firstly,the algorithm extracts the green component of the color images as gray image for further analysis.Then,the algorithm takes the local maximum points as the points of interest in the image.And56 dimensional LBP features in the near areas of the points will be extracted to train an ensemble classifier of RUSBoost.Finally,the points of interest will be divided into two kinds: positive predictions and negative predictions.(2)Extraction of regions of interest based on MSER.The regions of interest in the image will be extracted by the method of MSER which gets a series of fitting ellipses.Then,centers of the ellipses will be considered as the points of interest.And shape analysis will be used to select appropriate regions.(3)Hierarchical contour analysis algorithm.This chapter includes the illumination model,the deduction of annular illumination distribution.Firstly,multiple levels of contours around each valid ROI were extracted and fitted with the circular Hough transform.Then,multiple fitted circles would be merged into one target and be considered as the last detected citrus fruits.The algorithm based on LBP features and HCA based on a training set with 45 labeled color images will be evaluated with a test set of 20 images and reaches a recall rate of 77.9%,a precision rate of 76.4% and a F value of 77.1%.The average processing time of each image is 10.23 s.The algorithm based on MSER and HCA is evaluated with the same image sets and reaches a recall rate of 81.2%,a precision rate of 83.5%,and a F value of 82.3%.The average processing time of each image is 3.70 s.The research shows that the two methods can effectively identify the green citrus on trees.