Comparison study on convolution neural network (CNN) techniques for image classification / Siti Maisarah Zainorzuli ...[et al.]

• Main Authors: Zainorzuli, Siti Maisarah, Che Abdullah, Syahrul Afzal, Abidin, Husna Zainol, Ahmat Ruslan, Fazlina
• Format: Article
• Language: English
• Published: UiTM Press 2022
• Subjects: Neural networks (Computer science); Image processing
• Online Access: https://ir.uitm.edu.my/id/eprint/63165/1/63165.pdf; https://doi.org/10.24191/jeesr.v20i1.002; https://ir.uitm.edu.my/id/eprint/63165/; https://jeesr.uitm.edu.my/v1/; https://doi.org/10.24191/jeesr.v20i1.002

Deep Learning is an Artificial Intelligence (AI) function which can imitate the human brain to process data and deciding. It has networks that able to learn the unsupervised data that unlabeled or unstructured. It also identified as Deep Neural Network or Deep Neural Learning. Convolutional Neural Network (CNN) is a subset of Deep Neural Network which frequently used to analyse images. CNN also called as ConvNet which can be trained using an existing model that has been finetuned or trained from zero by using a large data set. CNN was often used in image classification due to its effectiveness and accuracy. However, there are several CNN architectures such as AlexNet, GoogleNet and ResNet-50. To select the appropriate architecture for our research in agriculture, a preliminary study to evaluate the architecture were conducted by using five different types of flower datasets that obtained from Matlab and Kaggle database. The three types of CNN architecture were compared in terms of accuracy in classifying the flowers. Result of this study indicated that the optimal configuration is by setting the number of epochs at 30, with the learning rate at 0.0005, to obtain the highest accuracy at 99.82%.