딥러닝 컬러 영상분석 기반 씨없는 수박종자 품종 선별

Watermelon growers face various problems in acquiring accurate seed germination and varietal mixing of 3x, 2x and 4x seeds, thus hampering the watermelon sector due to seed-ploid nomenclature confusion given the fact that seeds tend to look alike on attempt of the human eye. These circumstances indirectly inflict negative effects on the farmer income and the development of watermelon seed company. Therefore, seed purity is a prerequisite for all seed breeders and companies, as the performance of a given seed variety can be known and standardized from the rest. In this study we employed color camera techniques to discriminate triploid watermelon seeds from diploid and tetraploid seeds nondestructively. Seed-ploid images were acquired by both a digital canon and mvBlue fox3 USB3.0 camera and discrimination models constructed with multivariate machine learning methods of one class classification with DD-SIMCA and SVM quadratic. And to improve the accuracy obtained, a deep learning model was also developed. The models constructed on one class classification with DD-SIMCA, and SVM-quadratic method yielded triploid discrimination accuracy of 69.5% and 85.5% respectively when a digital canon camera was used. To further improve the class-ploid discrimination accuracy, a deep learning model of deeplabv3+ and Resnet18, gave an accuracy of 95.5%. Deep learning model results demonstrated a higher discrimination accuracy and thus these results can be potentially automated and applied on online system for real time seed-ploid discrimination and sorting.