An Improved Deep Neural Network for Classification of Plant Seedling Images

Abstract

This scientific pursuit aimed to develop a deep learning architecture tailored to classify plant seedling images. Our architecture encompasses seven learned layers - five convolutions and two fully connected. We performed full training on the network using 4, 234 plant seedling images belonging to twelve plant species from Aarhus University Signal Processing group. The system is fine-tuned for the architecture to have greater processing time and low memory consumption. The architecture was evaluated using different network parameters. Furthermore, we used training loss function, accuracy, sensitivity, and specificity to evaluate the system performance. Experimental results proved that the developed architecture has reached excellent performance with overall accuracy of 90.15%. Results were achieved in 111 minutes and 36 seconds. Future work includes, first, use the model with greater amount of datasets through data augmentation and compare the results to other existing deep learning architectures using same datasets. Second, authors will consider CNN and RNN architectures together using several other plant datasets. Third, create a portable mobile application for plant seedling images classification utilizing the developed model.