Corn Seedling Weed Identification In Machine Vision Research

Aim at alleviating the environment pollution which is caused by applying herbicide in weed protection, this paper used computer vision to identify weed from corn seedling, and to find out the position and growing state of weed. Accordingly it can provide theory and technology support for automation of spot spraying herbicide during corn seedling stage. The contents and results of this study are as follow.(1) According the property and demand of the study, the factors affecting system performance were analyzed in this paper. On the basis of Considering the ratio of performance and price, hardware which can meet the demand of performance and be extended were chosen to compose the computer vision system. Appropriate distance and view angle of camera were selected and linear dimension was calibrated. In order to make the weed identification system better adapt the various illumination intensity and soil humidity conditions, weed images were acquired in natural environment.(2)After statistical experiment on color indexes of weed images which were acquired in various illumination intensity, soil humidity and remains coverage, it pointed out that hue and relative color index are insensitive to illumination intensity, soil humidity, remains coverage and shadow, hue and excess green have a distinct contrast between plant and non-plant, but can not be used to identify corn and seed.(3) When the threshold of excess green, which is generated by red, green and blue without being normalized, is set at 30, it can correctly segment plant from soil background. The correctness is over 99%.(4) Modified hue can identify plant from non-plant background through Ostu's thresholding. The correctness is over 99%.(5) Shape features studied were aspect, first invariant central moment, elongatedness, roundness, circularity and thickness. Aspect and first invariant central moment are the most effective shape features for identifying monocotyledonous weed from dicotyledonous weed, and the correctness was 93%. Other features provide little information for classification. Thickness is an effective feature for identifying corn from monocotyledonous weed, and the correctness was 90%.(6) Six shape features were used to design back-propagation network for weed identification and the network structure was 6-12-3. The network classified over 90% of the weed from corn seedling. And the influence of learning error and the number of hide layer nodes on network performance was also studied.(7) Traditional Region labeling was not considered because it wastes time and scanfrequently the image. But fast edge tracing was improved to trace several edges each time. Chain code of the edge was used to calculate perimeter and object area. This method can increase the processing speed.(8) Software for weed identification was developed by use of Visual C++. It can be used in image analyzing and area of interest test. It can also provide software and technology support for developing weed identification system.