An Instance Segmentation Approach for Wheat Yellow Rust Disease Recognition

Abstract

Stripe rust is the most economically wheat rust disease. Wheat stripe rust has been reported in more than 80 countries and an estimated that 40% of wheat quality has been decreased due to wheat stripe rust disease. Detection of stripe rust diseases is very important for enhancing wheat quality production. Therefore, the detection of stripe rust disease is achieved either through experienced evaluators or computer vision techniques. The experienced field evaluator takes a lot of time for the detection of wheat stripe rust disease. Moreover, high cost is associated with experience evaluators. This paper proposed a mask-region based CNN (Mask-RCNN) that will identify the proper symptom location of stripe fungi on the wheat plant. The Mask-RCNN identifies the proper symptom location through object detection and semantic segmentation process. Initially, a total of 400 wheat images have been collected from primary and secondary sources. The VGG-16 and VGG-19 pretrained models are the backbones of the Mask - RCNN model for building and extracting features. After generating a mask of wheat stripe rust from full-size images, the Mask-RCNN predicts the stripe rust symptom location on the wheat plant. Different hyperparameters such as training rate, learning momentum batch rate have been used for Mask-RCNN training. During training time, the experimental result shows that Mask-RCNN achieves 0.9% mask_loss over 25,000 iterations. Throughout prediction for stripe rust disease symptoms in the wheat plant, the prediction rate in terms of F1-score of 90.63% and 87.25% for the diseased area of stripe rust and the wheat plant has been calculated respectively.