Automated Classification of Wrinkle Levels in Seed Coat Using Relevance Vector Machine

Seed-coat wrinkling in soybean is often observed when seeds are produced in adverse environmental conditions and it has been associated with low germinability. Manually rating seeds is time consuming, error prone and fatiguing – leading to even more errors. In this paper, an automated approach for the rating of seed-coat wrinkling using computer vision and machine learning algorithms is presented. The proposed system provides a GUI for ground truth annotation and a pipeline consisting of seed segmentation, feature extraction and classification using multi-class Relevance Vector Machines (mRVM). This research also proposes a reliable new feature for seed-coat rating based on texture. An additional contribution of this paper is a database of annotated seed images, which is being made available to researchers in the field. The results showed an accuracy in wrinkling rating of 86% for matches within ±1 scores from the ground truth.