Multi-level Banana Plant Disease Diagnosis with Score and Model Fusion

Many computer vision models designed for plant diagnosis predominantly focus on training with leaf images. However, real-world diagnosis often requires a comprehensive approach that extends beyond the examination of leaves alone. To address this limitation, we propose integrating information from both leaf and whole plant data using various score-based fusion schemes, aiming to enhance the reliability of disease identification. Utilizing a novel multi-view banana plant dataset encompassing four diagnosis classes, we compare conventional leaf-only or plant-only techniques with our proposed approach. We employ product, sum, and maximum score fusion methods to combine leaf and plant classification scores. The results indicate that adopting a combined leaf-and-plant approach improves diagnosis accuracy and sensitivity in comparison to single-view methods. Additionally, our findings suggest that combining scores from a joint classifier trained on both leaves and plants yields heightened sensitivity for specific classes compared to combining scores from distinct leaf and plant classifiers, each specialized in its domain.