Study On Green Crop Image Segmentation Based On Superpixel

Due to illumination,complicated background environment and other factors,there exist some areas of highlights and shadows,etc.on the surfaces of green crops,which decrease the ability to recognize the row lines of green crop for agricultural robot,so that it cannot complete the further extraction of guidance parameters of the vision guidance system.In this study,we focus on this issue,taking the real the green crops including corn,cabbage and broad beans as the research objects.Then we use different color index methods with thresholds and machine learning method to segment the green crops from their soil backgrounds under different illumination conditions.Finally,we evaluate the segmentation results by employing the objective evaluation indicators in this dissertation.The main contents and conclusions of this study are as follows:1)Green crop images acquisition and classification: Considering the common characteristics of green crop images under different illumination conditions,on the basis of existing algorithms,a crop image classification method based on gray histogram in combination with some statistical parameters such as mean,variance,skewness,and kurtosis is proposed in this dissertation.The experimental results show that the green crop images in this dissertation can be divided into two types: images under normal illumination conditions,images under abnormal illumination conditions,which include weak and strong illumination conditions;compared with the manual classification results,the MER(mean error rate)of classification of the proposed method is 3.30%.2)Green crop images segmentation under normal illumination conditions: because of color characteristics of green crop images is obvious under normal illumination conditions,the color index methods with adaptive thresholds,median filtering and morphological operations can be used for segmenting them with good effect.Due to the influence of complexity of the soil backgrounds on threshold selection,this dissertation subdivides those images into two categories,using different threshold to obtain the optimal segmentation results respectively.Finally,an objective image segmentation evaluation which is applied widely in this field is utilized to evaluate the performance of the algorithm proposed in this dissertation.Experimental results show that using different threshold methods for crop images under normal illumination and complex background with the same optimal color index method can obtain more plant details,and the average recall rate is increased by 5.56%.3)Green crop image segmentation under abnormal illumination conditions: since color characteristics of green crops images are disturbed to different degrees under abnormal illumination where the traditional color index methods with thresholds are not suitable for use in the experiments.Therefore,we introduce a correction term into the SLIC(simple linear iterative clustering)algorithm for image preprocessing to extract superpixels of the green crop images.Then we design an effective feature vector to train the CART decision tree in order to segment the green crops images from their soil backgrounds.Experimental results show that the areas of green crops in terms of superpixels are more accurately extracted by using the improved SLIC.Green crop segmentation is completely finished by CART based on these superpixels,more than the traditional color index methods with thresholds,even in which there are some high light or deep shadow areas.The average comprehensive score(F1)reaches 83.71% with a rise of 7.97% than the results of traditional color index methods with thresholds,indicating that the algorithms proposed in this dissertation can effectively segment green crop images from their soil backgrounds under abnormal illumination.Based on the above,the algorithms proposed in this dissertation could not only classify the green crops images in terms of the different illumination conditions,but also segment efficiently them,providing a reference method for agricultural robot in vision navigation or video image crop growth monitoring.