Study On Identification Of The Field Weed Based On BP Neural Networks

The threat of accompanying farmland weeds is the main reason for dispiriting growth trend and declining production of crop.The chemical method can prevent and control the weeds seedlings effectively and timely,which avoiding the crop yield affected by mature weeds.The chemical weeding is simple and time-saving for widespread vulgar spraying,yet there are much more disadvantages,such as water and soil pollution,pesticide residue,poisoning etc.which is contrary to the concept of green environmental protection,sustainable development,precision agriculture,etc.As the same time,fixed-point drilling is the main type of the modern crop planting,and weeds grow between ridge with clusters and randomness,so a fixed-point variable spraying system is needed to design.This research takes an example of common weed control problem of test field of corn of Jilin Agricultural University,and designs a weed identification system based on the combination of machine vision and image processing technology,which provides the basis for the development of real-time weeding equipment.The main research contents of this paper are as follows:Firstly,according to the analysis and comparison the emergence speed of weeds and crop and the requirement of image quality of the subsequent processing,we can determine acquisition time and the height and angle of the camera lens.Then,color features in RGB,2G-R-B,HIS,YCbCr color space are extracted to reach that 2G-R-B characteristic value can satisfy the requirement of gray level's change.Neighborhood average filtering and median filtering algorithm are used respectively to eliminate the noise and interference imported when image acquisition,transmission and transformation.Comparing the results we can found the effect of median filtering is better for field image with salt and pepper noise.Secondly,comparing three threshold segmentation methods and discovering that OTUS threshold method is the fastest,as the well,its' foreground area image is complete and non-noise but more noise background.Afterwards,mathematical morphological open operation is used for reprocessing and different size of corrosion cell and expansion unit are compared by the result,at the end,radius of 23 flat disc corrosion cell and diameter of 13 prismatic expansion unit are selected.Thirdly,marking the connected component of each independent blade getting from morphological segmentation,then it's easy to calculate the area and moment features parameters based on the regional characteristics.Then,five kinds of edge detection method are used for it and the test result of canny operator is the most ideal.After that,calculating the dimensional parameters based on contour feature,for example,circumference,length and width,etc.By comparing the characteristic parameters of corn,weed and amaranth the three plants,it is found that the width/height ratio,circular degree and the first invariant moment can distinguish plant species effectively which are appropriately as input characteristic parameters of weeds classifier.Finally,weeds classifier is established based on BP artificial neural network for it has the unique function of storage,imagination and recognize for complex field image.Through experiment design,we can gain the optimum combination of the number of hidden layers,learning rate and momentum factor.After the above,we can make a conclusion that the network structure is MLP: 361 and the above factors are 6,0.5and 0.5 in order.The BP neural network was established based on the MATLAB simulation,and with a total of 120 set of characteristic parameters of crop and weed field images as the input of the recognition analysis.Among them,90 set as the training samples,30 groups as the test samples,and the recognition accuracy of simulation were 98.89% and 93.33% respectively.