An adaptive image fusion-based data augmentation method for plant diseases based on GANs

Current datasets in the field of plant disease target detection suffer from issues such as insufficient sample size and discrepancies between sample images and actual field environments. Data augmentation based on image fusion can combine target images and background images to increase sample size, but when the shooting environments of target images and background images differ significantly, it can result in prominent boundaries and mismatches between target colors and background colors in the fused images, causing the generated sample images to deviate from actual field environments. To address these issues, this study proposes an adaptive image fusion-based data augmentation method for plant disease detection based on Generative Adversarial Networks (GANs). First, the discriminator is trained to determine whether the environmental features of the target and background in the fused images are consistent. Then, the generator is trained to learn how to generate fused images with consistent environmental features by offseting the features of the target image. Finally, the trained generator is used to perform adaptive image fusion enhancement on the plant disease dataset. Experimental results from plant disease target detection on the original real dataset show: The mean Average Precision (mAP) of the target detection model YOLOv9s trained using the adaptive image fusion-based data augmentation method proposed in this study is $88.1 \%$, which is higher than the mean Average Precision of target detection models trained using unenhanced data sets ($84.9 \%$), the Random Erasing method ($87.1 \%$), the Mosaic method ($87.5 \%$), and the image fusion method ($86.7 \%$), effectively improving the detection performance of the target detection model for plant diseases.