Lightweight Convolutional Neural Network Model for Cassava Leaf Diseases Classification

The world population has increased many folds in the last few years and crossed the figure of 7 billion. Africa has the highest population growth rate. Food and water are the first and foremost necessities for the survival of living beings. Cassava is among the staple foods of Africa and other countries. Its roots and leaves fulfill the daily caloric demands of millions of people. In the last few years, the production of cassava has decreased due to the spread of disease in cassava leaves. The manual identification of these diseases needs large number of people with sufficient level of skills in this field, so an automated technique is required that can help farmers by detecting cassava leaf diseases within a small time frame. The paper proposed a lightweight convolutional neural network (CNN) model to find four types of cassava leaf diseases i.e., cassava mosaic disease, cassava green mottle, cassava bacterial blight, and cassava brown streak leaf disease. It uses depthwise separable convolution operation to reduce the number of parameters that makes it suitable for mobile devices. The proposed model has used both channel attention and spatial attention to highlight disease part of the leaf and suppress background scenes. The modified attention mechanism of the proposed work helps in further improving the accuracy of the model. The experimental results exhibit that the proposed model outperforms other state-of-the-art models like VGG16, ResNet 50, EfficientNet, MobileNet V1, and MobileNetV2. The proposed model uses 1.1 million fewer parameters than MobileNet V2 that makes it suitable for low computing power devices. It uses the natural scene images for the classification that enhances its applicability in real-world scenarios.