A Method for Data Augmentation of Diseased Leaf Detection

Efficient and accurate detection of plant diseases and timely taking protective measures are important measures to ensure the yield and quality of agricultural products. However, training an accurate detection model requires a large and diverse dataset to ensure that the trained model has a high accuracy rate in a variety of environments. In order to solve the problems of overfitting diagnosis and low quality of diseased leaves, a data augmentation model PlantGAN based on improved CycleGAN was proposed. In this paper, four kinds of grape disease leaves were selected from Plantvillage data set for experiments. An attention network was added to each generator in the CycleGAN network to transform related regions (leaf regions). A new labelless leaf segmentation module, LeafSeg, was proposed. So that the segmentation module can be more accurate, without changing the surrounding environment of the leaf parts to cut and convert. The results showed that the average accuracy of the network trained with the original dataset without data enhancement was 87% in the new environment, and the average accuracy of the model trained with the contrast network CycleGAN after data enhancement was 89.5% in the new environment, which was 2.5% higher than the baseline, and the effect was not significant. The average accuracy of the model after data enhancement using the PlantGAN network proposed in this experiment is 97% in the new environment, which is 10% higher than the baseline. Moreover, the disease transformation image generated by this experimental model is of higher quality than the image with data enhancement. The disease transformation images generated by the PlantGAN model proposed in this experiment can not change the surrounding environment, and the generated images have clear disease details and high quality. This method can effectively solve some serious problems of overfitting in the final disease detection process caused by fewer data sets of disease leaf causes and single environment.