The Research Of Identifying Algorithm And Image Features' Optimization In Weed Identification

After Weed identification divided to four stages as image gain, segmentation, identifying and following process, image feature and algorithm are optimized with the character of weed image to improve weed identification in nature and provide theory for automatism weed control system. The main contents are as follow:(1) In segmentation whose aim is distinguishing vegetable and background, red weeds' color features are studied before seeding when vegetable only included weeds without plant. For the image needs identification to distinguishing plant and weeds, a new algorithm based on 2-dimensional Histogram has been proposed to solve the distortion since illumination. The algorithm could be potentially useful for improving identification because of saving the leaf connection. Firstly the color feature and thresholding for weed image segmentation is transformed into the segmentation surfaces in color space after analyzing the probability of vegetable and background. Secondly to optimize color feature and thrsholding the segmentation surfaces parameters' selection is processed via genetic algorithm that is based on evaluating segmentation error with Bayes formula. A comparison of the segmentation results between Excess-Green that is most popular color index in weed segmentation and new segmentation surface -149R+218G-73B by the genetic algorithm was conducted. The results show that the segmentation error ratio compared theory optimization reduced from 2.47 times to 1.47 times in Nightshade and weed image segmentation. In cotton fields weed image segmentation, the precision of Excess-green and G are 84.69% and 75.01%, and new feature increase to 91.67%.(2) In identification for distinguishing plant and weed, four shape features including Aspect, Elongatedness, Roundness and Compactness are used to identify green capsicum and Goosegrass, Iran Speedwell, Sunn Euphorbia with fuzzy. On images with small quantities of overlapping, vegetal species are well recognized. To select identification feature, the mathematics model are process via analysis of identification precision and process time. The results show the combination of Aspect and Compactness is best. On images with much quantities of overlapping, the new algorithm based on the ratio of skeletionization length and leaves area for identifying cotton field weed image. It was found that segmentation precision doesn't concern identification precision deeply and the results of different segmentation features all can get good precision in identifying. If segment with new segmentation feature -149R+218G-73B, the plant and weed precision are 82.77% and 59.69% and the existing features can't distinguish them effectively. The new methods can have a good effect on identifying weeds from plants while there are serious overlap leaves because of depending on whole plant.(3) In following process for spraying parameters the genetic algorithm is used to optimize after building mathematics model of spraying efficiency. The spraying parameters of new method improve the spraying efficiency, and the improving efficiency is more prominent for complex weed image.The identification efficiency is improved via optimising image feature and evolving algorithm for automatism weed control equipment. A novel algorithm is from image gain to processing spraying parameters after analysing the character of weed identification, and the practicality of weed identification with image process is improved.