Automatic disease symptoms segmentation optimized for dissimilarity feature extraction in digital photographs of plant leaves

Abstract

Segmentation of diseased symptom regions in images of plant leaves is a crucial stage in the application of machine learning for plant diseases detection. This process also known as Region of Interest (ROI) segmentation involves separating purely color variant symptom lesions from surrounding green tissue from which discriminant features are later extracted. However, investigations have shown that vivid anatomy of a disease symptom progression right from inception to manifestation through which finer disease characterization dissimilarity features can be fostered are not captured in a segmented ROI. Furthermore, the typical ROI segmentation process is often plagued by challenges ranging from intrinsic factors such as image capture conditions to extrinsic factors such as disease anatomy where symptoms fade into healthy green tissue the separation boundary to become impalpable. This adds further complexity to the process or produce erroneous result. This research proposes an automatic extended region of interest (EROI) segmentation to incorporate symptom progression information by extending the border region to cover some part of healthy tissue using color homogeneity thresholding. To produce a ground truth, the typical ROI segmentation alongside a reduced ROI were implemented on a well-known PlantVillage dataset from which separate textural and color features were extracted and used to build a linear classifier. A comparison between the classification results further reinforced the advantages of the proposed approach for dissimilarity features extraction. Through this research, finer characterization features can be extracted for the classification and severity estimation of plant diseases.