Improved Line Detection in Images using Neural Networks and DTE Subclassifiers

It is widely accepted that deep neural networks are very efficient for object detection in images. They reach their limit when multiple long line instances have to be detected in very high resolution images. In this paper, we present an original methodology for the recognition of vine lines in high resolution aerial images. The process consists in combining a neural network with a subclassifier. We first compare a traditional U-Net architecture with a U-Net architecture designed for precision agriculture. We then highlight the significant improvement in vine line detection when a DTE is added after the customized U-Net. This methodology addresses the complex task of dissociating vine lines from other agricultural objects. The trained model is not sensitive to the orientation of the lines. Therefore, our experiments have improved the precision by around 15% compared to our improved neural network.