Transfer Learning for Diploid Haploid Corn Seeds Image Classification using Residual Network Wahyudi Setiawan, Haris Muhajir Al Fatih, Yoga Dwitya Pramudita
Department of Information System, University of Trunojoyo Madura, Jawa Timur
Abstract
The corn seed breeding program is the initial stage of getting quality plants. It also aims to increase corn growth production and productivity. Actually, corn seeds have two types: diploid and haploid. Almost all corn seeds are naturally diploid, and less than one percent are haploid. However, haploid seeds can be the forerunner of double haploid (DH) corn seeds produced by biotechnological engineering. DH seed just need two to three generations of breeding, in contrast to manual crossbreeding, which can take up to eight generations. During the corn harvest, diploid and haploid seeds mix. Carrying out the separation manually requires more effort and needs to be improved by visual limitations and human energy. In this study, we classified diploid and haploid corn seeds. Data uses public sources with 3,000 images consisting of 1,230 haploid and 1,770 diploid images. This research uses Residual Network 50 with transfer learning. Initialization of hyperparameter values consists of learning rate 0.001 & 0.0001, epoch 64, batch size 64, and Stochastic Gradient Descent optimizer. In the training process, k-fold cross-validation is used with a value of k=5 to divide the train data and validation data. This research has a test scenario: two different models (resnet-50 with and without transfer learning). Based on testing results, using the resnet-50 with transfer learning has an accuracy 95.55%, precision 97.12%, recall 95.29%, and f1-score 96.19%.
Keywords: Corn Seeds, Image, Classification, Residual Network, Transfer Learning
