Extraction method for centerlines of rice seedlings based on SUSAN corner

Abstract: In south China, the rice seedlings present various morphological characteristics during the growth period. What's worse, duckweed and cyanobacteria, whose colors are very similar with the rice seedlings, appear in the paddy field frequently. The complicated environment makes it challenging to extract the guidance lines in south China. Domestic and foreign scholars have proposed many methods to detect the guidance lines. But most of them are difficult to be applied in paddy fields in south China. In order to solve these problems, a new method which is based on SUSAN (smallest univalue segment assimilating nucleus) corner and nearest neighbor clustering algorithm is presented. The method consists of 4 main processes: image segmentation, feature points detection, feature point cluster and guidance lines extraction. Firstly, the color image is transformed into grey scale image using normalized ExG (excess green index). In this process, the distribution area of the crops can be extracted from the background. But there is a lot of noise in the grey scale image after this process. Secondly, SUSAN corner algorithm is used to detect the feature points in the grey scale image. The target crop regions were obtained by detecting the feature points. And most of the noise in the grey scale image can be filtered. In order to make the SUSAN algorithm adaptive, we propose an equation to compute the corner threshold. Thirdly, feature points are clustered using nearest neighbor clustering algorithm. There are 2 steps to cluster the feature points. Accordingly in the initial step, the image is scanned by a scanning window and then the feature points are clustered preliminarily. After that, the feature point groups are clustered in vertical direction. The center point clusters of each target region were obtained by using the clustering algorithm. Finally, the known point Hough transform is applied in the algorithm in order to extract the center line of each cluster rapidly and effectively. In order to test the algorithm, 3 growth stages are taken into consideration. The circumstances of 3 growth stages are different from each other. The significant differences of the 3 growth stages are: in the first growth stage, there are few duckweeds in the water; in the second growth stage, there are a lot of duckweeds in the water; in the third growth stage, there are a lot of cyanobacteria in the water and the crops are close to each other. Then 3 image datasets are used to test the algorithm. The images of the datasets are taken in a paddy field in South China Agricultural University. The test result shows that the highest accuracy rates are 87%, 89% and 85% respectively in the first, second and third growth stage. It also shows that the runtime of the algorithm is 352 ms in the first growth stage, 405 ms in the second growth stage and 563 ms in the third growth stage. The results indicate that not only the algorithm is able to detect the guidance lines accurately but also the run time of this algorithm is acceptable.