Comparison of Pre-trained and Convolutional Neural Networks for Classification of Jackfruit Artocarpus integer and Artocarpus heterophyllus

Cempedak (Artocarpus heterophyllus) and nangka (Artocarpus integer) are highly similar in their external appearance and are difficult to recognize visually by a human. It is also common to name both jackfruits. Computer vision and deep convolutional neural networks (DCNN) can provide an excellent solution to recognize the fruits. Although several studies have demonstrated the application of DCNN and transfer learning on fruits recognition system, previous studies did not solve two crucial problems; classification of fruit until species level, and comparison of pre-trained CNN in transfer learning. In this study, we aim to construct a recognition system for cempedak and nangka, and compare the performance of proposed DCNN architecture and transfer learning by five pre-trained CNNs. We also compared the performance of optimizers and three levels of epoch on the performance of the model. In general, transfer learning with a pre-trained VGG16 neural network provides higher performance for the dataset; the dataset performed better with an optimizer of SGD, compared with ADAM.