Research On Farmland Images Segmentation Under Different Illumination

Due to weather,temperature,illumination and other factors,there will be uncertainties in the obtained information when intelligent agricultural robots perceive the environmental information.It is necessary to segment the farmland images under different illumination conditions,in order to further improve the environmental awareness ability of intelligent agricultural robots.In this study,the farmland images,which were belong to the three-leaf stage to the five-leaf stage in Northwest Agricultural and Forestry University experimental field,were as research objects.Different color index,threshold segmentation methods and machine learning methods were adopted to realize the segmentation of farmland images under different illumination conditions.Subjective evaluation methods and objective evaluation methods were applied for algorithm analysis and verification.The main contents and conclusion of this paper are as follows:(1)Farmland images acquisition and classification.In order to realize the automatic classification of farmland images,a method based on mathematical statistics to analyze the histogram of farmland images was proposed.The experiments results show that the mean and skewness indicators about the histogram corresponding to the R,G and B color channels of the farmland images are not coincident in any interval.Therefore,they can be used as the standard for automatic classification.Compared with artificial classification method,the experiments results show that the maximum error rate of the proposed method is 10.52%,which can be used for farmland images automatic classification.(2)Farmland images segmentation under sufficient and weak illumination conditions.Since the color characteristics of the farmland images was obvious under sufficient illumination conditions,the histogram mean method and Otsu method were used for farmland images segmentation.Similarly,because the color and shape characteristics of the farmland images were unobvious under the weak illumination conditions,fuzzy C-means clustering algorithm(FCM)based on unsupervised learning was used to implement the segmentation of farmland images.Finally,the experimental results were analyzed and compared.Because the segmentation targets under the sufficient and weak illumination conditions were relatively complicated,then subjective evaluation method was used to analyze the experimental results.The experiments results show that the average subjective quality fraction of the two types ofimage segmentation results are 4.26 and 4.06 respectively.And,according to the five grading quality scale and obstacle scale of CCIR500,the image segmentation quality based on the method in this paper is better.In this way the segmentation of complex farmland images can be realized.(3)Farmland images segmentation under strong illumination conditions.In this paper,because large number of high light points under strong illumination conditions would have negative impacts on image segmentation precision,the improved simple linear iterative clustering algorithm(SLIC)was used for preprocessing to obtain the super-pixel.After that,we extracted feature vector,drew the color feature curves and selected the optimal feature vector through the curve.Then the classifier was established to realize the farmland images classification.The classifier contained the Bayesian method and support vector machine(SVM).The experiments results show that the improved SLIC can reduce the running time without affecting the preprocessing results.The overall classification accuracy of SVM is better than Bayesian method,and the average overall classification accuracy can reach94.83%,which indicating that SVM can effectively realize segmentation of simple farmland images within large number of high light points.Based on automatic classification of farmland images,this paper fundamentally realized the segmentation of farmland images under different illumination conditions,which provides a powerful guarantee for improving the ability of intelligent agricultural robots perceiving the environmental information.