QBNet: A Lightweight Quantized Model for Dynamic Banana Disease Recognition

Bananas are an extremely in-demand and widely consumed fruit in Bangladesh. Due to their easy availability and low cost, bananas are a favored fruit among people of all ages in our country. Banana diseases are capable of causing significant damage to banana cultivations across the world. These diseases put banana production at risk, causing crop damage and financial losses for producers. Farmers can save plants from disease by identifying them early and improving their quality and quantity. Many researchers have introduced new approaches that combine deep learning and machine learning to detect banana plant leaf disease. These approaches require a lot of computational power and time. The extensive hardware requirements of complex architectures, such as CNNs, limit the effective execution of resource-constrained devices. In this study, we introduce a lightweight quantized model, QBNet, for easy deployment in resource-constrained devices. The proposed QBNet is the quantized version of the BNet model. BNet is a deep neural architecture-based model that can distinguish between healthy leaves and those affected by three distinct diseases. Our QBNet model, employing int8 quantization, achieves a 12-fold reduction in weight compared to BNet, resulting in a compact model size of only 1.2 MB. The experimental results show that the QBNet model outperformed several transfer learning models, namely DenseNet201, ResNet50, and VGG19, with an accuracy of 99% while the BNet model achieved an accuracy of 99.21%.