Automated Wheat Disease Classification Under Controlled and Uncontrolled Image Acquisition

This paper presents a practical classification system for recognising diseased wheat leaves and consists of a number of components. Pre-processing is performed to adjust the orientation of the primary leaf in the image using a Fourier Transform. A Wavelet Transform is then applied to partially remove low frequency information or background in the image. Subsequently, the diseased regions of the primary leaf are segmented out as blobs using Otsu’s thresholding. The disease blobs are normalised and then radially partitioned into sub-regions (using a Radial Pyramid) representing radial development of many diseases. Finally, global features are computed for different pyramid layers and combined to create a feature descriptor for training a linear SVM classifier. The system is evaluated by classifying three types of wheat leaf disease: nondiseased, Yellow Rust and Septoria. The classification accuracies are slightly over 95% and 79% for images captured under controlled and uncontrolled conditions, respectively.