Segmentation of Green Vegetation in the Field Using Deep Neural Networks

Green vegetation segmentation in the wild from digital images is an essential step in the precision agriculture. Most of the existing methods apply plant characteristics or handcrafted features to segment plants and backgrounds. In this paper, we propose a semantic segmentation method based on deep convolutional neural networks to extract vegetation from complex and natural field conditions. Pixel-to-pixel prediction is performed with multiscale features and abundant local information, which are fully extracted from the designed network architecture. Different vegetation indexes are exploited as the inputs to train the neural network on a typical crop dataset. Experiment results show that the presented approach achieves an effective and comparable performance on segmentation of green vegetation. Besides, the end-to-end deep semantic vegetation learning framework is fast enough to realize real-time segmentation, which is available for agriculture automation in the fields.