Tea Leaf Disease Classification Using an Encoder-Decoder Convolutional Neural Network with Skip Connections

Abstract

In this study, a Convolutional Neural Network (CNN) model for tea leaf disease classification is introduced, which utilizes an encoder–decoder structure with skip connections. The proposed model is designed to reduce computational complexity and generate high-resolution images from lower-resolution representations obtained using the encoder. The skip connections in the model enable the transfer of information between the encoder and decoder, thereby allowing for the reconstruction of the original dimension image from the encoded representation. The performance of the proposed CNN model is compared against several popular computer vision models, including GoogleNet, ResNet50, and VGG16. An ensemble model is also developed to further improve classification performance. In addition, two hybrid CNN models are created, one by combining the computer vision models without using them as feature extractors, and the other by using their feature extraction capabilities with various machine learning algorithms for image classification. Experimental results indicate that the proposed CNN model with skip connections outperforms the individual models with accuracy, precision, recall, and F1-score 87.36, 88.12, 87.36, 87.30, respectively. Based on our experimental results, our proposed CNN model with skip connections did not outperform the ensemble and hybrid models in terms of accuracy. However, the difference in accuracy was marginal, and our proposed model offered lower computational complexity, making it a more efficient option. Therefore, while our proposed model may not be the best performer in terms of accuracy, it can still provide a viable alternative for tea leaf disease classification, especially when computational efficiency is a critical factor.