The Research Of Image Processing And Identification Method For Crop’s Kernels

Detecting the appearance characters and identifying the breed for the crop’s grain withcomputer vision is important to improve the agricultural automation, increase theeffectiveness of agricultural production and strengthen the competitive advantage of Chineseagriculture. Based on these reasons, this thesis selects the static images of the kernels of rice,corn and so on as the research objects. The research designs and develops a system based oncomputer vision which could detect the kernels’ appearance characters and recognize them,and the research emphasis includes the image segmentation technology, the extraction ofmorphological features and the identification methods. The major work is as follow:(1) Extracting the kernels of corn and rice from the background by using the iterativethreshold method and Filling the hole in the kernel. On these base，some segmentationmethods for the occluded kernels are studied.(2) In order to solve the problem of the over-segmentation of the rice kernel induced byseveral local minima after distance transform by watershed algorithm, an improved methodbased on the result of distance transform is proposed. The proposed method merges theadjacent local minima points into one region by morphological dilation operator, which makeseach rice kernel only has one local minimum as possible. Then watershed algorithm segmentsthe preprocessed image. The experimental results show that the proposed algorithm getsrespectively accurate segmentation ratio of87%,93and92%and89%for four types of rice.(3) A segmentation method based on active contour model is proposed. The regionboundary of local minima after distance transform is taken as the initial curve. Under theguidance of active contour model, the curve converges to the edge of the kernel, and eachkernel in the images is segmented finally. The segmentation accuracy for long glutinous rice,round glutinous rice, non-glutinous rice and black rice reached88.0%,93.4%,92.4%and90.4%respectively.(4) In order to detect the broken kernels and identify the kernels’ breeds, a set ofmorphological parameters of kernels are extracted to measure their size, shape and color.Besides, according to the national standard of rice, the average lengths of long glutinous rice，round glutinous rice, non-glutinous rice, red aromatic rice, Thailand aromatic rice and blackrice are estimated.(5) The tipcap is an important appearance character of many crop kernels. A tipcapidentification algorithm based on Harris corner detector is developed. Harris algorithm detectsthe corner within a local region. As the most obvious corner structure, the tipcap has the maximal response of Harris corner detector, which makes it can also detect the kernelswithout the obvious tipcap accurately. The kernels with tipcap of corn, pumpkin and summersquash are selected for test. The experimental results on750kernels show that the proposedmethod gets overall accuracy of95.6%for tipcap of these kernels.(6) An algorithm of departing the embryo part and non-embryo part of the corn kernelis proposed. Firstly, on the base of location of the tipcap of the corn kernel, a triangle regionextended from the tipcap is used to approximate the fruit stalk, and this region is deletedconsequently. Then linear combinations of single color channel R+B-G and G+B-R are usedto change the color image of multi-channels to two new feature image, and their binaryimages are get with the iterative threshold method. The embryo part of corn kernel isextracted from intersection set of the two binary images after morphological optimizing.(7) On the base of distinguishing analysis of whole kernels and broken kernels, theparameters of area, eccentricity, circularity, the proportion between the longer axis and theshorter and three invariant moments are selected as characters to detect broken rice by usingSVM. The detected accuracies about long glutinous rice, round glurinous rice andnon-glutinous rice are100%,94%and92%respectively. The broken corn kernels are detectedbased on SVM, and the overall accuracy is92.5%.(8) An identification method based on sparse representation was proposed fordiscriminating varieties of rice and corn precisely. The rice images of six varieties and thecorn images of four varietes were taken as the research objects. To represent single rice kernel,its color and morphological characters were extracted. All training samples make up the datadictionary of the sparse representation. For each variety, the projection of the testing sampleon the data dictionary is calculated. The breed, which had the minimum projection error,would be regarded as the right kind of rice which the testing sample belonged to.Experimental results demonstrated that the overall identification accuracy of the proposedalgorithm for the six rice breeds expressed with13features and the four corn breedsexpressed with16features were99.6%and88%respectively.(9) A system based on machine vision is designed and developed by using MATLABlanguage, which could detect the kernels’ appearance quality and classify them. The softwares of this system is composed of file managing, image preprocessing, segmenting theoccluding kernels, tipcap recognizing, extracting the embryo part of the corn kernel, breedidentification and so on.