A Convolutional Fusion Architecture for Processing Smart Phone based Images to Detect Banana Leaf Diseases

Abstract

Banana diseases cause significant loss to the cultivators and it is essential to automate the diagnostic process to prevent inconsistencies and delays in disease detection by human experts. This study proposes a convolutional fusion architecture called RAttnNet that integrates separable convolution in traditional convolution with attention mechanism within a unified architecture to enhance feature extraction while utilizing support vector machine (SVM) for better classification accuracy. Although in this field research is going on, most of the earlier works relied on publicly available datasets and computational complexity of these models presents opportunity for improvement in both accuracy and computational costs. A dataset of banana leaves is created for this study by capturing banana leaves images, with smartphones, from different parts of Assam in India. The performance of RAttnNet is evaluated with seven other state-of-the-art models and it revealed that RAttnNet outperformed others achieving an impressive accuracy of 99%. This study showcases the potential of deep learning tools in agriculture and encourages further research and application in this domain.