Abstract:
Many countries around the world depends on agriculture, as it helps reduce poverty,
increase national income, and improve food security. However, plant diseases often
impact food crops, leading to significant annual losses and economic setbacks
in agriculture. The best solution of the problem is to identify the plant disease
as soon as possible so that necessary steps can take. Traditionally, humans have
identified plant diseases visually, but this method is often slow, and also the number
of domain experts are less. Recently, there has been significant progress in using
deep learning to classify plant diseases. However, the main problem is to collect the
sufficient annotated image data to train these models effectively for plant disease
classification. Also the limited training data can negatively affect the performance
of CNN models. To address this, we designed a Deep Convolutional Generative
Adversarial Network (DCGAN) to overcome the issues of over-fitting
and to increase the dataset sizes. Here we worked on the dataset called DiaMOS
Plant dataset, consisting of 3006 images of pear leaves of four classes (3 diseases
and one healthy class). The dataset was very imbalanced, so we used DCGAN on
the minority classes separately to enhance the dataset. We developed some CNN
models for classification and compared with some Pre-trained models (VGG16,
ResNet50, Inception V3). The results showed an average increment of classification
accuracy.
