The Innovation Of Machine Vision Target Identification Method For Horticultural Application

In horticulture, there are many important tasks which mainly rely on visual identification and determination. The automated and intelligent machine vision could be used to do the tasks instead of manual vision so as to reduce the labor intensity and improve the efficiency of management. With the rapid development of artificial intelligence technology, the main content of this paper was the research and innovation the methods of identification horticultural object based on machine vision, mainly in the following two aspects:firstly,2DFT spectrum analysis method was innovatively using to solve the pest counting and leaf disease severity of quantitative analysis; Secondly, in order to improve the yield prediction accuracy, it was proposed that the fruit tree’s canopy characteristics such as the number/size of fruit, the dense degree of canopy and the situation of fruit clusters were extracted from the digital image. Based on these characteristics, the yield prediction model was established by artificial intelligent method.1) A novel smart vision method based on the two dimensional Fourier transformation (2DFT) spectra In this study was presented. Rather than directly counting the pests captured on the traps, the novel idea is to consider the trapped pests as a type of noise in a2D image and2DFT to be a specific noise collector. The experiment included quantifying analogue pests with two different distributions/patterns on templates, whiteflies (Bemisia tabaci) on yellow sticky traps (YSTs) and thrips (Frankliniella. occidentalis) on blue stick traps (BYTs). The proof-of principle test with a range of analogue pests on four templates (R2=1) verified that the2DFT-based index is solely dependent on the pest numbers captured on the traps and independent of the pest distributions/patterns on the traps. The experimental results (R2=0.9994and R2=0.9989) from the whiteflies on YSTs and the trips on BSTs highlighted the proposed method. In addition, the measurement errors were also addressed relating to some possible situations. The relative error was within4%.2) An automatic measuring method of leaf area based on2DFT spectrum analysis method was proposed. This method was used to determination the complicated shape leaves of tomato and simple shape leaves of sunflower. The agreements were both above0.99between the2DFT counting method and manual counting method, and the relative errors were in±7%. Monitoring experiment with continuous changes of leaf area were carried out, sunflowers, pepper and eggplant plants as the object. The correlation between the three kinds of plant area of the vertical projection obtained by2DFT spectrum analysis method and manual measurement of total leaf area were above0.91. In addition, in the case of tomato leaf mold disease, the quantitative identification of leaf disease severity was achieved based on2DFT.3) A yield estimation method based on fruit tree’s canopy characteristics was proposed.230of "Gala" apple trees were selected as research objects.165of them were for half-ripe date when the most apples in tree just turned red and ripe date when apples were ripe. Four features were extracted which were the proportion of fruit area, the proportion of fruit number, the proportion of small size fruit area and the proportion of leaf and fruit. Then two yield prediction models were established using back-propagation neural network (BPNN) based on these features. One was for the tree with half-ripe apples, the other was for the tree with ripe apples. Two estimation models for testing sets, R2was more than0.87between the estimation of yield and the actual yield.65sample trees were for the very early yield prediction just after June drop. Five features were extracted which were the proportion of fruit area, the proportion of fruit number, the proportion of leaf and fruit, the proportion of partly shaded fruit number and the proportion of partly shaded fruit area. The estimation model for testing set, R2was more than0.68between the estimation of yield and the actual yield. In addition, the "Pinova" apple tree was selected as object to study the influence of carrying out different crop load management (CLM) to fruit recognition.