Classification of Haploid and Diploid Maize Seeds based on Pre-Trained Convolutional Neural Networks

Abstract

Analysis of agricultural products is an important area that is widely emphasized today. In this context, with the development of technology, computer-aided analysis systems are also being developed. In this study, a system has been proposed for classifying maize seeds as haploid and diploid using pre-trained convolutional neural networks. For this purpose, AlexNet, GoogLeNet, ResNet-18, ResNet-50, and VGG-16 pre-trained models have been used as feature extractors for the haploid and diploid seed classification process. In the first stage, the deep features of haploid and diploid maize seeds have been obtained in these models. The features have been taken from different layers of network architecture. Instead of softmax classifier in the last layer of the network, classifiers based on decision tree, k-nearest neighbor, and support vector machine have been used. According to the classification results with these features, the achievements in network architectures and classifier methods have been observed. The experiments have been carried out on a publicly available dataset consisting of 3000 haploid and diploid maize seed images. The experimental results revealed that the developed classification systems demonstrate a remarkable performance.

Hyperparameter Optimization in Convolutional Neural Networks for Maize Seed Classification

“…even though if a model is started with relatively wrong parameters at the beginning of the training, as the number of epochs increases, it is possible to catch up with the best models. Our best validation accuracy is %94.66 which is higher than best validation scores of [2] and [6]. Finally, the total number of parameters belonging to the best model we created was 9,022,299.…”

“…Thanks to the classification of images by CNN, costs can be greatly reduced by eliminating the need for specialists. CNNs have also been the subject of previous articles such as [4][5] [6] in the process of distinguishing haploid kernels from diploid kernels. The main differences in the studies in this area are due to the depth of the CNNs used, the activation functions, and the hyperparameters of the neural networks.…”