Research On The Image Segmentation Method Oriented To Weed Identification

Ecological agriculture is a new trend of agricultural development, and weeding technology is gradually developing to precision variable spraying. Realization of precision variable spraying relies on high accuracy of weed identification. While edges extracted by traditional edge extraction methods exist some problems, such as non-continuity, incompleteness, incline, jitter and notches etc, and these methods are sensitive to noises. Moreover, K value of classification algorithms is uncertain. Therefore, this paper studied weed identification in maize field from two aspects: firstly proposed an edge extraction method based on graph theory to solve imprecise problem of edge extraction; secondly, based on that extracted features were reduced and selected by principle component analysis, and two classification algorithms with different K value were used to carry out experiments, then determined appropriate K value. Main work and conclusion were as follows:(1)Determination of image preprocessing and threshold segmentation method. Leaves of maize and green glaucum were used as an example, methods were determined in every phase, that is, ultra-green method was used for gray processing; median filtering was used for filtering process;OTSU method was used for image segmentation.(2)Proposed an edge extraction method based on graph theory. The method included three phases: in pixels similarity calculation phase, the weights were given to sides in undirected graph that was constructed, which represented pixels similarity; in threshold determination phase, the mean of all the weights(the similarity of the whole image) was determined as threshold, and it was a global threshold and can represent all the information of image and overcome the subjectivity of directly given threshold; in edge determination phase, when weights on horizontal or vertical sides were smaller than the threshold, the left nodes of horizontal sides and the up nodes of vertical sides were retained to constitute edges of image. Experiments without and with noises were carry out to verify the performance of this method by comparing it with Roberts operator, Sobel operator,Prewitt operator, Lo G operator and Canny operator. Experiments show that this method can overcome deficiencies of five operators above, such as non-continuity, incompleteness, incline,jitter and notches etc., and its anti-noise capability is better than five operators. Finally, it was used to precisely extract edges of weeds and maize.(3) Extraction and selection of feature. 15 feature parameters were extracted from three aspects including color, shape and texture of maize, endive, green glaucum, metapress japonica,solanum nigrum and Calystegia hederacea. These 15 feature parameters were reduced from 15 to10 by principle component analysis. Finally, 10 features were used for weed identification.(4) Weed identification experiments. Both LM algorithm and KNCN algorithm with different K value(K=1~8) were used for color feature identification(CFR), shape feature identification(SFR) and texture feature identification(TFR) and comprehensive feature identification experiments(CSFR). Experiments show that with the increase of K value, total identification rate of CSTFR is obvious higher than that of CFR, SFR and TFR, and when K value is 1, total identification rate of CFR, SFR, TFR and CSTFR are the highest, which are 85.56%,86.67%, 86.67%, 88.89% respectively. Finally, KNCN algorithm with K=1 is used for weed identification based on the principle that higher total identification rate of CSTFR and shorter running time.The research can provide certain reference value for weed identification in maize field and theoretical support for precise variable spraying, meanwhile as technical reserve for its implementation.