MoSViT: A Lightweight Vision Transformer Framework for Efficient Disease Detection via Precision Attention Mechanism

元奇 陈 ^* 爱萍 王 子扬 刘 杰 岳 Zhang Enxu 飞 李 宁 张 ^*

• Xijing University, Xi'an, China

Maize, a staple food crop globally, faces substantial yield reductions due to diseases. Traditional diagnostic methods are often inefficient and subjective, leading to significant challenges in timely and accurate pest management. To address these issues, this paper proposes an innovative classification model, MoSViT, incorporating advanced machine learning and computer vision technologies. Building on the MobileViT V2 framework, we integrate the CLA focus mechanism, DRB module, MoSViT Block, and the LeakyRelu6 activation function to enhance feature extraction accuracy while reducing the complexity of analyzing all possible features. Trained on a comprehensive dataset of 3,850 images covering Blight, Common Rust, Gray Leaf Spot, and Healthy conditions, the MoSViT model demonstrates remarkable performance, achieving classification accuracy, Precision, Recall, and F1 Score of 98.75%, 98.73%, 98.72%, and 98.72%, respectively. These metrics surpass those of other leading models, including Swin Transformer V2, Densnet121, and EfficientNet V2, not only in accuracy but also in model parameter efficiency. Moreover, the model's interpretability is significantly improved through the use of heat maps to analyze the focus areas, providing valuable insights into the decision-making process. Testing on small sample datasets further confirms the model's excellent generalization capability and its potential for enhancing small sample detection scenarios.