Research On Image Segmentation Of Green Crops In Natural Lighting

With development of modern information technologies,agricultural modernization and information has become a main trend in China’s.Image processing as a prerequisite is a basis of machine vision.And its importance is more self-evident.This thesis mainly focuses on the research on field crop image segmentation.Under daily lighting conditions,there are different reflections and shadows in field crop image.This thesis is deployed mainly around the two issues.In the first step,the information of each pixel is extracted in the crop and background regions of the original image,in terms of R,G,B with R/G,G/B,B/R,ExG,ExGR,CIVE,VEG,COM,respectively.The principal components analysis of the eleven variables is performed to obtain the front two principal components with the maximal eigenvalues.The two principal components can be used to represent green crop or soil background.The first seven variables of ExG,ExGR,VEG,COM,R,G,and B are selected according to their weighting coefficient sizes as an input vector of Elman neural network.Experimental results show that such an input vector is more effective for green crop segmentation compared with the traditional indices.In the second step,for some slightly reflective images of green crops,the R,G,and B color components are converted into Lab space to segment them by using neural network.At the same time,the morphological operations of expansion and erosion are used to eliminateisolated pixels or small areas to contribute to segmentation evaluation.In the third step,there are still some crop images with strong reflect light or severely shadow regions that cannot be segmented completely in Lab space.Therefore,the texture structure features of the image are considered to solve the two problems,which include the differential texture average,contrast,second moment,and entropy.These features characteristics are analyzed based on the reflective or shadow regions in image.Among them,the differential texture contrast has better ability to discriminate between shadow or reflective region and normal region in crop image.Therefore,the L,a,b and the difference texture contrast feature in Lab color space are used as an input vector of Elman network to segment green crop image from the soil backgrounds.Experimental results show that this segmentation method is effective for green crop image with strong light or severe shadows.