GRAPE LEAF DISEASE RECOGNITION BASED ON A MULTISCALE MIXED ATTENTION RESIDUAL NETWORK

A multiscale efficient channel attention spatial-residual network (MECAS-ResNet) is proposed in this paper. On the basis of ResNet50, the training speed and model volume are reduced by group convolution. Through multiscale convolution, the spatial attention and effective channel attention are fused into the model to improve the network's attention toward various regions of lesions. A Wasserstein generative adversarial network is used for data enrichment, and the resulting images are diverse and variable. The enriched dataset is fed into the MECAS-ResNet model for identification and comparison to other models. After the improvement, the optimal recognition accuracy reached 95.31%, which is 10.02% higher than that on the original dataset; the model size is only 30.88 MB; and the average F1-score reached 95.30%. The improved model has higher recognition performance for similar diseases in different degrees, which is better than other models. The data generated by the experiment can be used for grape disease recognition, which greatly reduces the cost of data collection and provides technical support for grape disease recognition and precise drug treatment.