Practical Tobacco Diseases and Pests Recognition Based on Gans and Improved Resnet

Precise recognition of pests and diseases during the tobacco growth stage is crucial for green pest control, directly affecting tobacco yield and quality. To address the challenges of insufficient image quality, limited dataset size, and low accuracy in tobacco disease and pest recognition, this study proposes a method based on generative adversarial networks and an improved residual network. First, Leaf-GAN was employed to augment tobacco disease images, leveraging its label-free leaf segmentation module to generate high-quality synthetic samples. Real-ESRGAN was then applied to enhance the resolution of pest images. The residual network was further optimized by integrating an Efficient Multi-Scale Attention module and Ghost Convolution, improving both feature extraction and computational efficiency. Additionally, Class-Balanced Focal Loss was introduced to mitigate class imbalance. Experimental results show that the proposed model achieves 95.04 % recognition accuracy with an inference speed of 7.20 ms per image and only 8.01 MB parameters, outperforming existing methods in balancing accuracy and computational cost.