论文部分内容阅读
In Korea, the one which has the highest productivity among the upland crops is soybean.The productivity of soybean is the second largest after that of rice.For soybean production, the major problem is weed control.Soybean production is mostly affected by negative influence of weed.According to the agricultural technology guide book for soybean from Rural Development Administration in Korea, Weeds cause decrease of productivity, physiological troubles, and loss of work efficiency.If weeds are not controlled, soybean yields can be decreased from 32% to 77%.In the case of the United States, the loss of soybean yield occurred by weed amounts to about 1.9 billion dollars every year.Therefore, this experiment was conducted to extract vegetation from plant image and to detect weed in soybean field, as a basic study on development of weed control system using machine vision technology for precision agriculture.The plant used in this experiment is soybean (Glycine max (L.) Merr.).This was cultivated from June until November in 2014 at the college farm of Gyeongsang National University.25 experimental images were acquired in term of a week from late June to early August, using an RGB digital camera (D7000, Nikon Corporation,Japan) with 18-55mm zoom lens.The image resolution was 986 pixels by 653 pixels.In order to extract vegetation from background, we used two vegetation color indices for generating contrast image, which were (1.4R-G) from RGB and (b-a) from Lab color space.Then a combination of automatic threshold methods, Otsu's and Triangle threshold, were used to create binary image.The results were evaluated by comparing with the other method based on K-means clustering approach and manually segmented reference images.Weed detection algorithm was performed using the binary image generated by our method.It is based on morphological difference between the plants.This algorithm detects boundaries in the binary image, and calculates some parameters such as Area and Area/Perimeter.If the calculated values not satisfy the condition, the detected region is filled with red color as a weed.The algorithm of those procedures was built by MATLAB software (R2012a, MathWork, USA).This tool is a framework approach that most necessary functions for image processing task are basically provided.We additionally coded some function which is not provided using M-script.As a result of comparison analysis, there was no significant difference (p=0.756) between the accuracy of our segmentation method and that of the K-means clustering method.Moreover, the processing time of our method was about twice shorter than that of the K-means method (p<0.001***).Thus we found that our proposed method is more efficient than the K-means method.As a result of weed detection, the accuracy was 80% and it is confined to the image containing low density of weed.This is because the weed detection algorithm is not allowed to detect weeds when the image contains extremely dense weed population.Thus we are endeavoring to solve this limitation using a pattern recognition approach.