Weed identification from corn seedling based on computer vision

Computer vision and artificial neural network were used to identify weed from corn seedling. First, the Otsu's method for automatic threshold was applied to segment weed images based on the modified excess green feature to distinguish the plant objects from the background. Second, successive erosion and dilation were implemented on the binary image. Third, the rest objects were classified into corn and weed by back-propagation neural network according to their shape features: aspect ratio, circularity and first invariant moment. Finally, all the weed objects were obtained by erasing corn objects in the segmentation results using seed filling method. The results showed that the algorithm for weed identification based on back-propagation neural network correctly classified 87.5% corn objects and 93.0% weeds. The average processing time was about 58 ms for a 640×480 pixels weed image.