Recognition of perennial herb images in uncontrolled environment using enhanced convolutional neural network

Automated medicinal plant recognition using a robust deep learning architecture is necessary to classify various types of medicinal plants accurately. Nevertheless, developing a deep learning architecture is very challenging by the lack of publicly available medicinal plant image datasets especially for uncontrolled environments. Hence, the research proposes a new architecture with less Convolutional Neural Network layers that can identify medicinal local perennial herbs under uncontrolled environment. To perform the experiment, an image dataset consisting of 34200 images of twelve different medicinal local perennial herbs in Malaysia was developed. The images were captured in various imaging settings, including varying scales, illuminations, and perspectives which adds diversity in terms of interclass and intraclass within the dataset. The dataset can be accessed at https://www.mylpherbs.com/. This dataset will motivate researchers to use cutting-edge computer vision techniques to categorize or create models for plant identification in the presence of background noise and obstructions. Hence, by utilising this developed image dataset, a Convolutional Neural Network architecture was constructed. The efficiency of EfficientNetB0, EfficientNetB7, and the proposed architecture in classifying the developed image dataset was investigated using batch sizes of 32, 64 and 128 and steps per epoch of 50 and 80. The performance of the architectures in classifying different amounts of images was also investigated. The results were analysed using ANOVA One-Way Analysis and performance metrics. The findings demonstrated that the proposed architecture produced 98.33% training accuracy and 91.99% validation accuracy, outperformed EfficientNetB0 and EfficientNetB7 when experimented using 128 batch size and 80 steps per epoch. The proposed architecture demonstrated an improvement of 15.02% for training accuracy and 1.59% for validation accuracy when compared to EfficientNetB7. Meanwhile, EfficientNetB0 only achieved 92.38% when trained using the same hyper-parameters. Furthermore, the proposed architecture has a computation time of 44.53 minutes, lower than EfficientNetB0 of 143.97 minutes and EfficientNetB7 of 841.41 minutes. As for dataset size, the ANOVA One-Way Analysis showed that the number of images used in the experiment does have a significant impact on the accuracy achieved. The proposed architecture achieved the lowest training accuracy of 98.02% and validation accuracy of 67.58% when using the Quarter dataset size as the low number of images denotes fewer data learned by the architecture. However, increasing the number of epochs increases the training accuracy to 99.88% and validation accuracy of 83.44% without compromising the computation time. The proposed architecture is able to identify the images outperformed both EfficientNetB0 and EfficientNetB7 in terms of accuracy and computational time.

Text FK 2023 28 - Declaration Form.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (454kB) Text FK 2023 28 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (3MB) Text FK 2023 28.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (3MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18944

Actions (login required)

View Item