Semantic Segmentation of Wheat Stripe Rust Images Using Deep Learning

Wheat stripe rust-damaged leaves present challenges to automatic disease index calculation, including high similarity between spores and spots, difficulty in distinguishing edge contours. In actual field applications, investigators rely on the naked eye judge extent, which is subjective, of low accuracy, essentially qualitative. To address above issues, this study undertook a task semantic segmentation wheat rust damage images using deep learning. problem small available datasets, first large-scale open dataset from Qinghai province was constructed through greenhouse image acquisition, screening, filtering, manual annotation. There were 33,238 our with size 512 × pixels. A new paradigm defined. Dividing indistinguishable spots into different classes, accurate background, leaf (containing spots), investigated. assign weights high- low-frequency features, we used Octave-UNet model that replaces original convolutional operation octave convolution U-Net model. The obtained best benchmark results among four models (PSPNet, DeepLabv3, U-Net, Octave-UNet), mean intersection over union 83.44%, pixel accuracy 94.58%, 96.06%, respectively. showed state-of-art can better represent discern information region improve spores, leaves, backgrounds dataset.