Few-Shot Learning for Plant Disease Detection using DeepBDC

Plant disease detection in low-data scenarios is a major challenge for computer vision applications in agriculture. Traditional deep learning needs big, labeled datasets, but these are often not available for rare or new plant diseases. To address this issue, this paper presents a few-shot learning (FSL) method for classifying plant diseases using an improved Deep Brownian Distance Covariance (DeepBDC) framework. The model uses a ResNet-12 as a backbone network which uses a Convolutional Block Attention Module (CBAM) to help the network focus on important spatial and channel-wise features. In network, dropout regularization is also applied to reduce overfitting, which is common in few-shot tasks. The DeepBDC module is improved by using L2 normalization and temperature scaling, which makes the feature representations more stable and better for classification. The model uses cosine similarity for 1-shot learning and Gaussian kernel similarity for 5-shot and 10-shot settings. The method is evaluated on two public plant disease datasets, PlantVillage and CCMT, using standard 5-way classification with 15 queries per class and 2000 episodes for each configuration. On the PlantVillage dataset, the model gets 46.53% accuracy for 1-shot, 65.86% for 5-shot, and 69.67% for 10-shot tasks. On the CCMT dataset, it reaches 40.19%, 49.19%, and 53.67% accuracy for the same settings. These results show that the proposed method is a useful way to detect plant diseases when there is less data.