Weed Identity Based On Computer Vision Technology

As is known, agriculture is very important in China, but the problem about weed has hampered the further development of Chinese agriculture. The chemical weedicide can weed with high efficiency while the extensive spraying causes high cost and the destruction of entironment. Variable rate spraying was presented against these disadvantages. The key technique of achieving variable rate spraying was studied in this paper, it is named recognizing weed technique. The main content and achievements were presented as follows:1. Summarized the recent development of the study at home and abroad on recognizing weed by computer vision technology, pointed out the problems of the these study and a comparatively comprehensive expatiation on computer vision system was presented.2. Build up and perfected the computer vision system for this study. The system was consisted of 3CCD multi-spectral imaging sensor, image acquisition card and computer.3. Total 114 images of weed and soybean which belong to 38 groups in field were gained. Each group has the image of three color component as red, green and near infrared (NIR).4. New explore was carried out to do the background segment by using color characteristics. In order to find out the most efficient color component for background segment, 16 new color characteristic components were developed based on R, G and IR. Then r, g and ir were gained by calculating R, G and IR and other 15 components were also gained continually. Three steps were executed. First, the images of 37 color component mentioned above were calculated using Matlab7.0.1 software. Second, the histograms of these images were calculated and compared based on Matlab software. 9 color component images with double peaks in the histogram image were divided into two groups: r, r+g, r+ir, r+g+ir, r-ir and G, IR, ir, g+ir. The last step is to calculatethe global threshold T of these 9 color component images of two groups. Thethreshold T will be used for separating the background and weed. The result showed that the color component of ir, r+g and r+g+ir performed well.5. Further research about ir, r+g and r+g+ir was executed using Matlab software. After the background segment and the de-noise which was achieved by using median filter processing for the 3 color component images about ir, r+g and r+g+ir, the gray images were further executed to eliminate the particular details of the images. By comparing the result of the operation of open and close for these 3 components. The color component of ir was the best for close operation while the other two color component (r+g and r+g+ir) would do a better performance on open operation. Finally, the results of these 3 color components for recognizing weed showed that ir was the best. So, as a conclusion, the color component of ir is best for recognizing weed based on computer vision. Also the operation process for recognizing weed was proposed.6. Supposed an idea on the variable spray mechanism of weedicide.