Field Weeds Identification Based On Multi-spectral Images

Weed automatic identification technology is a key prerequisite for variable spraytechnology. There is often light, shelter and other factors in the physical environment of thefield which affect obtaining characteristics parameters of the plant leaves. Also there are stillpoor accuracy and low efficiency in weed identification method. To solve these problems andimprove the accuracy and efficiency of weed identification, this thesis takes weeds in cornfield as the research object and study the soil background segmentation method based on themulti-spectral image fusion. After separating plant leaves and extract shape, texture andfractal dimension three features of leaves, this thesis study three weed recognition ofmulti-feature combination based on C4.5algorithm, multi-feature fusion on the SVM-DS,PCA and SVM weed identification algorithm. The work and conclusions for the thesis are asfollows:(1)According to the existed problems for the weed identification at home and abroad andthe actual needs of field weeding, a field multi-spectral image acquisition platform wasdeveloped which consists of a mobile car, multi-spectral vision system, portable computer,and power systems. The platform can real-time acquire field images of weeds and crop toexplore new way for automatic weeding from the actual condition of the field.(2)Taking thistle, chenopodium album and Convolvulus arvensis three common weeds incorn field as research object, using MS4100multi-spectral camera in field image acquisitionplatform to acquire multi-spectral weed images; by comparing the segmentation results ofIR-R, G+IR-R, IR-R-GR, IR./R four fusion methods in different lighting conditions,different soil moisture and ground residue cover to proven IR-R fusion with Otsu thresholdsegmentation method can effectively separate the plants from the soil and other fieldbackground.(3) Extracting method of plant leaf shape features, texture features and the fractaldimension for weed identification was studied. Six kinds of shape feature parameter like therectangularity, elongation ratio of width to length, density，circularity and first invariantcentral moment of shapes were extracted from the leaves blade; using gray or gradient featureon the basis of gray-gradient co-occurrence matrix to describe the inside and edge informationof blade images separately, extract the small gradient strengths, gray-scale heterogeneity,energy, correlation and inertia etc. five texture feature parameters that can effectively distinguish weeds; blanket algorithm was used to extract the fractal dimension of the leaves.(4) Multi-feature combination weed identification method based on the algorithm ofC4.5was proposed and it is simple with high accuracy; using charactes like the shape featureparameters, texture feature parameters, fractal dimension as input to build weed identificationmodel can can effective support decision-making of the weak classification (singlecharacteristic) to strong classification (multiple combinations of features), and reduce thewrong identification caused by the similarity samples in single-feature recognition to someextent. The test results showed that the combination of the multi-feature, shape, texture,fractal dimension combinations has best recognition rate by97.22%, higher than the twofeature fusion recognition rate among texture features, shape feature and fractal dimension.(5) Studied and proposed weed identification method based on SVM-DS multi-featurefusion. Took full advantage of the superiority of SVM in solving the small sampleclassification problems and multi-feature information fusion ability of DS evidence theory ineffectively reduce effect on the recognition result caused by conflict features. Firstly, using"one to one" multiclass SVM to give basic belief assignment (BPA) of evidence (weedcharacteristics) in the same discernment frame, then using fusion algorithm based on matrixanalysis DS to fusion these data and obtain the final recognition result according to thedecision-making rules. The test results showed that the average accuracy rate of multi-featurefusion was96.11%with less volatile; the accuracy and stability were significantly higher thansingle feature recognition.(6)Weed recognition algorithm based on PCA and SVM was studied and proposed. Theoriginal12dimensional feawture data was reduced to6dimensions by using PCA; thecharacteristics with a greater contribution to the classification rate were retained; therebyreducing the input dimension of classifier and learning complexity and weed recognitionbecame more quickly and efficiently. Experimental results showed that the recognition ratebefore and after dimensionality reduction was95.83%and97.22%respectively;time-consuming for identification was0.0102s and0.0056s respectively. It showed that it isentirely feasible to idnety the weeds by using the feature parameters reduced by PCA insteadof using the original feature parameters. This algorithm significantly reduced the time forweed identification algorithm at the same time of guarantee the recognition rate.(7)The thesis compared the pros and cons for the three kinds of weed identificationmethod of C4.5algorithm multi-feature, the multi-feature fusion method of SVM-DS, methodof PCA and SVM. The former two methods were found to be simple, but both can realize theuse of feature information by a linear combination or change the use of multi-featureinformation, and more dependent on the acquisition of the characteristics data and information. Meanwhile, they are lack of stability due to the influence of conflict orcontradiction characteristics; Due to the DS fusion algorithm, the SVM-DS multi-featurefusion focus more on the mutual influence or support function between feature information, itcan effectively reduce the impact of individual abnormal characteristics, with good stabilityand accuracy. The deficiency is that the algorithm is slightly more complex and BPA access ismore difficult.