Recognition of Plants by Leaves Digital Image and Neural Network





Recognition of plants by leaves digital image and neural network

To identify different plants by leaves digitalimage is one key problem in precision farming. By thecombination of image processing and neural network,Most of the image blocks of different plants could becorrectly classified. Firstly, the image enhancementprocessing can make objects in the source image clear.Secondly, due to the different shapes and sizes of imageblocks of leaves, they could be separated and extractedfrom sources. Then, by using image analysis tools fromMatlab, these characters such as sizes, radius, perimeters,solidity, and eccentricity could be calculated. Then, usingthem as input data, create a radial basis function neuralnetworks. Divide the input data into two parts. Select onepart to train the network and the other to check thevalidity of the model. Finally, input data from otherimage frames under the same condition could be used tocheck the model. In this work, the total accuracy is about80%.These methods was simple and highly effective, Sothey could easily be integrated into auto machines in thefield, which can largely saving labor and enhanceproductivity.

INTRODUCTIONTo control weeds mechanically in crop field, and tosave fertilizer and to protect environment, Themachines in the field need to see where is crop andwhere is weed. So, how to recognize them automatedis in great need. Many works have addressed thisproblem with various solutions. Some authors haveworked with infrared images instead of color imageswhich using only the shifts in gray level forsegmentation of plants in the images[1]. Infraredimages may also be suited for methods that do notrequire segmentation[2]. More recently, fixed thresholdvalue for segmentation of infrared images was used inmany systems. Most processing methods usedsegmentation for plant identification. A kind ofprocessing method without segmentation step wasdeveloped[3], which has been proven can reduce thecomputational burden of the image processingsoftware. The machine vision system now a days isnormally consists of computer, digital camera andapplication software. Many kind of algorithms areintegrated in the application software. Image analysisis one important method that helps segment image intoobjects and background. One of the key steps in imageanalysis is feature detection. Segmentation algorithmsfall into two general classes, based on whether theysearching for discontinuities or similarities. Algorithmsfocusing on locating discontinuities in the data areprimarily edge-based, while algorithms concerned withlocating adjacent pixels based on similarities areprimarily region-based. Threshold techniques, a majorcategory of algorithms, can fall into either class. Inaddition to these two major classes, there are also anumber of general subcategories. For instance,algorithms either process color or gray-scale data,operate on either an individual pixel basis (global) or aneighborhood of pixels (local), and may use differentwindow sizes or different color representations[4]. Forexamples of surveys of segmentation algorithms,see[5]. Cheng discussed the major segmentationapproaches for segmenting monochrome images:histogram threshold, characteristic feature clustering,edge detection, region-based methods, fuzzytechniques, neural networks, etc[6]. He also reviewedsome major color representation methods and theiradvantages/ disadvantages. Ismail Kavdır applied aneural network classifier to differentiate between 2 and3 weeks old sunflower plants and common cockleburweeds of similar size, shape and color[7]. Most of thereported systems did not optimize the wavelength

Free download research paper







FREE IEEE PAPER