Research On Identification For In-field Weed/Corn Seedlings By Digital Image Processing

To maintain high crop yield, weed control is essential. Nowadays, herbicides are typically applied uniformly throughout the entire field in China. The blanket herbicide treatment results a series of environmental contamination. To solve this question, the study mainly focused on identification for young corn/weed seedlings in the field by digital image processing and pattern recognition, is greatly significant to perform selective spraying technology in the field, which means using the right dose rates of herbicides at the right place according to the distribution of the weeds, to save cost and protect the eco-environment. On the basis of analyzing relevant researches at home and abroad, this work presented in this dissertation is devoted to segment the vegetation objects in-field from different backgrounds under outdoor natural lighting conditions and identify the corn/weed seedlings, which provides feasibility study for real-time system to identify corn/weed seedlings. The main content on this dissertation are as follows:(1) Set-up of image acquisition system, study objects and in-field environment in this work were investigated and analyzed. According to trial objects and desired results obtained from this experiment, a machine vision system was designed and developed for taking images with an appropriate resolution and fixed camera height above the ground in corn fields. Moreover, a digital illuminance meter was applied to represent in illumination variations in this work.(2) A series of image processing foundational algorithms, such as image graying, image denoising, image segmentation, edge detection operators and morphology calculus, were analyzed in this study. Corn/weed identification system in this research was configured in these ways by selecting different image processing algorithms in each stage.(3) Considering actual in-field conditions, color vegetation indices in RGB space were proposed to separate plants from soil and residue background images with an automatic threshold. Moreover, an accuracy algorithm of detecting plant was described for the first time in this dissertation. Analyzing color vegetation indices of samples in image set, it was found that the ExG-ExR+0 method, which provided the best accuracy of 96.57%, was be less sensitive to lighting variations, and had the potential to work well for different residues and soil moisture backgrounds.(4) Width to length, roundness, elongation and seven invariable moments were selected as shape features for discriminating between the corn and the weeds. A comparison study of the classification capacibility of SVM and BP neural network models was conducted. It was found that the SVM classifier provided the best classification performance and was capable of classification accuracy of 95%, which exceeded the BP classifier accuracy of 90%. Furthermore, the SVM classifier also provided the same classification accuracy, up to 95%, when only using seven invariable moments as input features. In this experiment, the weeds overlapped the corn were handled and separated using morphology calculus.(5) Contrast, energy, homogeneity, correlation, mean, standard deviation, smoothness, third moment, uniformity and entropy were generated as texture features for discriminating between the corn and the weeds. A comparison study of the classification capacibility of SVM and BP neural network models was conducted. The classification results showed that plant species with the SVM classifier were successfully identified with correct classification rate of 90%, while the BP classifier only gave a correct identification of 80%.(6) Energy values were calculated from wavelet coefficients by using two-level wavelet decomposed to gray level images. Then the obtained seven energy parameters were used as input vector to construct BP network classifier. The results showed that monocotyledon and dicotyledon could be totally separated with 100% accuracy, whereas weed/corn seedlings could not be effectively separated only with correct classification rate of 77.14%.(7) Fractal dimension of corn/weed seedlings as feature parameter was first used in this work. To extract feature variables, image segmentation was done using H channel in HSI space. Three computational formulas of fractal dimension were proposed and compared in this experiment. The results showed that the Bouligand-Minkowski method was better than the other methods, and mean fractal dimension of corn seedlings and weed seedlings was equal to 1.204 and 1.079, respectively. The method based on SVM classifier was capable of identification accuracies of 80%.(8) Corn/weed detection based on multi-features for the purpose of improving recognition rates based on single feature was presented. First PCA was applied to reduce the dimensionality of the feature dataset in which there were a large number of interrelated variables, in order to obtain a new dataset of feature variables that were uncorrelated. Then BP classifier was used to show the effect of identifying corn/weed seedlings in this experiment.(9) Images were simulated under dynamic condition, and their restoration images using different parameters were estimated by qualify functions. In this experiment, the weeds overlapped the corn seedlings were first handled and separated using region labeling method under dynamic condition.