LightFormerMSC: a lightweight transformer with multiscale convolution for plant disease recognition

This paper presents LightFormerMSC, a lightweight Transformer-based architecture for plant disease recognition under complex field conditions. The model integrates inverted residual blocks with depthwise separable convolutions and SE attention to enhance local feature extraction while reducing computational cost. A multi-head self-attention module is employed to capture long-range dependencies and global contextual relationships. To address the morphological scale variation and background complexity of field disease patterns, a multi-scale feature fusion module aggregates texture features via parallel dilated convolutions and adaptive channel-wise reweighting. A normalized cosine margin classifier is adopted to improve inter-class separability under open-set conditions. Experimental results on benchmark datasets show that LightFormerMSC achieves a favorable balance between accuracy and efficiency compared to 13 representative lightweight models. It achieves 97.33% accuracy with only 0.36M parameters, making it suitable for deployment on edge devices in precision agriculture scenarios.