Corn And Weed Seedlings Detection Based On Multi-spectral Images

Vulgar herbicide spraying without distinguishing crops and weeds would not only giverise to the waste of herbicide and labor forces, but also cause the breakage of crop quality andenvironmental pollution, seriously affected the sustainable development of agriculture. Weedautomatic identification technology is the key prerequisite for variable spray technology todevelop intelligent weed control equipment to realize automatic variable spray. This work,taked summer sowing corn and associated weeds in the field of Northwest A&F university asthe main research object, acquired their multispectral static images, researched field weedsegmentation in the actual background and overlap or adhesion situation, compared corn fieldweed identification methods combined feature extracting and pattern recognition, whichprovide theoretical basis and technical support for the development of real-time field weedidentification system. The main contents of the thesis are as follows:(1)According to the existed problems for the weed identification at home and abroad andthe actual needs of field weeding, a mobile car platform with MS4100multi-spectral cameraand some equipment were used to acquire different background conditions (different lightintensity, soil moisture, stubble coverage) plant images with the right shooting angle, heightand resolution et al.(2)Synthesizing each channel information from multispectral image, taking relativesegmentation error and intuitive segmentation effect as the measure, comparing thesegmentation results of IR-R, G+IR-R, IR/R three fusion methods in different lighting intensity,soil moisture and ground stubble coverage, which proving that the IR-R fusion with Otsuthreshold segmentation method could effectively separate the plants from the soil and otherfield background. In view of the actual conditions, overlap and adhesion may existed in plantleaves, then making use of the related morphology operation and marker controlled watershedto separate mild and seriously overlapping leaves, which ensured the accuracy of feature data.(3) Extracting method of plant leaf shape features, texture features and the fractaldimension for weed identification was studied. Six shape features such as rectangularity,elongation ratio of width to length, density，circularity and first invariant central moment wereextracted from the leaves blade; using gray-gradient co-occurrence matrix to describe the inside and edge information of blade images separately, then the small gradient strengths,gray-scale heterogeneity, energy, correlation and inertia etc were extracted; blanket algorithmwas used to extract the fractal dimension of the leaves. On this basis, making use of PCA fordimensionality reduction to acquire the ahead5main principal components.(4) Weed identification with multi-feature combination based on the PCA_Soft setalgorithm was proposed. By using the advantage of Soft set that it had unlimited conditionparameter settings, taking characters such as shape, texture and fractal dimension as inputfeature vector to build weed identification model. In order to verify the validity of the model,leading weed identification algorithm based on Na ve Bayes, BP net and SVM into experimentto contrast the weed recognition efficiency and time. The results showed that the Soft setperformed best under the condition of considering recognition efficiency and real-timeperformance, the average recognition rate and time were96.56%and0.17s. Putting the5mainahead vectors taking from PCA operation as the input vector to weed recognition based on Softset theory, the result showed that the average recognition rate and time were97.13%and0.091s, which demonstrated the PCA_Soft set is feasible.