The Identification Research Of Wheat Breed Based On Computer Vision System

Wheat is one of the most common food crops in China, having a whole lot to do withsocial stability and economic development. Using computer visual technique for wheatquality detecting and classification has the advantages of intactness, high speed, and highaccuracy. It can emancipate the labour, and increase the working efficiency.This research will work on six un-classified wheat kinds（L8998、Neixiang188、9023、Youzhan1、Yumai47、Zhoumai12）. BENQ₅000E scanner is adopted for the acquisition ofmultiple wheat images. Then the colour （R、G、B、H、S、I）, shape （perimeter、square、rotunditydegree、rectangle degree、elongation degree） and texture （energy、entropy、contrast、inertiamoment、local stationarity） of the seeds’ features are extracted separately by image processingtechnology.During the identification stage, this research will try to introduce the texture of the wheatseeds into the classification system, in allusion to the problem of the traditional wheatclassification system happened when the un-classified seed kinds are multiple. First, theidentification effect which is only based on the colour and shape features is explored.Identification can be achieved via building the BP neural network. Then input the extractedtexture features combined with the colour and shape features, and rebuild the neural network.At the end of this research, genetic algorithm is used to optimize the network, avoiding thebreak of the network training while the local optimum is achieved. Furthermore, the MIValgorithm is used to calculate the average result of each input feature, and the influence ofeach feature on the classification result is investigated.According to the experience result, when relating to the identification of three categoriesof wheat, the input data of texture features combined with the colour and shape featuresmakes a better show than that of the latter two alone, with accuracy jumping from92.5%to94.17%. Therefore the texture feature is proved to be the hard evidence of the seedclassification. As far as six categories identification is concerned, the accuracy of optimizedneural network rise about3%from the original network.