Multi-Scale Attention Based Plant Disease Segmentation Network

In this paper, we propose a novel approach to automate the detection of plant diseases in field images captured by unmanned aerial vehicles (UAVs). In general, UAVs images have complex backgrounds, making it difficult to accurately identify disease regions on plants. To overcome this challenging problem, we proposed a multi-scale attention-based segmentation network that automatically extracts both plant leaves and disease regions based on the characteristics of each class. The network consists of two branches that independently segment plant and disease regions using a transformer-based self-attention mechanism and a cross-attention module for feature fusion. The proposed method outperforms existing fusion methods and improves segmentation accuracy.