Modified U-Net for the Segmentation of Sunflower Leaf from Photographs

Abstract

The paper presents an algorithm for segmenting sunflower leaves infected with diseases such as downy mildew and leaf scars. For the experiment, real images of infected sunflowers from the fields were used. The training set was generated manually. The modified U-Net used was able to achieve 87.52% accuracy on a training set of 150 image-mask pairs. The Dice score coefficient reached a value of 0.6900 after training the neural network for 80 epochs. The resulting masks completely replicate the contours of both the leaf itself in the foreground and the contours of the lesions, as well as leaf fragments in the background of the image. Such an algorithm can become part of a disease recognition system in precision farming tasks, increasing the accuracy and speed of the neural network for recognizing sunflower diseases. This will enable a prompt response to the source of the disease and prevent the infection of the entire field.