Abstract:

Objective Rice diseases significantly impact both the yield and quality of rice production. Automatic recognition of rice diseases using computer vision is crucial for ensuring high yields, quality, and efficiency. However, rice disease image recognition faces challenges such as limited availability of datasets, insufficient sample sizes, and imbalanced sample distributions across different disease categories. To address these challenges, a data augmentation method for rice leaf disease images was proposed based on an improved CycleGAN model in this reseach which aimed to expand disease image datasets by generating disease features, thereby alleviating the burden of collecting real disease data and providing more comprehensive and diverse data to support automatic rice disease recognition. Methods The proposed approach built upon the CycleGAN framework, with a key modification being the integration of a convolutional block attention module (CBAM) into the generator's residual module. This enhancement strengthened the network's ability to extract both local key features and global contextual information pertaining to rice disease-affected areas. The model increased its sensitivity to small-scale disease targets and subtle variations between healthy and diseased domains. This design effectively mitigated the potential loss of critical feature information during the image generation process, ensuring higher fidelity in the resulting images. Additionally, skip connections were introduced between the residual modules and the CBAM. These connections facilitate improved information flow between different layers of the network, addressing common issues such as gradient vanishing during the training of deep networks. Furthermore, a perception similarity loss function, designed to align with the human visual system, was incorporated into the overall loss function. This addition enabled the deep learning model to more accurately measure perceptual differences between the generated images and real images, thereby guiding the network towards producing higher-quality samples. This adjustment also helped to reduce visual artifacts and excessive smoothing, while concurrently improving the stability of the model during the training process. To comprehensively evaluate the quality of the rice disease images generated by the proposed model and to assess its impact on disease recognition performance, both subjective and objective evaluation metrics were utilized. These included user perception evaluation (UPE), structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and the performance of disease recognition within object detection frameworks. Comparative experiments were conducted across multiple GAN models, enabling a thorough assessment of the proposed model's performance in generating rice disease images. Additionally, different attention mechanisms, including efficient channel attention (ECA), coordinate attention (CA), and CBAM, were individually embedded into the generator's residual module. These variations allowed for a detailed comparison of the effects of different attention mechanisms on network performance and the visual quality of the generated images. Ablation studies were further performed to validate the effectiveness of the CBAM residual module and the perception similarity loss function in the network's overall architecture. Based on the generated rice disease samples, transfer learning experiments were conducted using various object detection models. By comparing the performance of these models before and after transfer learning, the effectiveness of the generated disease image data in enhancing the performance of object detection models was empirically verified. Results and Discussions The rice disease images generated by the improved CycleGAN model surpassed those produced by other GAN variants in terms of image detail clarity and the prominence of disease-specific features. In terms of objective quality metrics, the proposed model exhibited a 3.15% improvement in SSIM and an 8.19% enhancement in PSNR compared to the original CycleGAN model, underscoring its significant advantage in structural similarity and signal-to-noise ratio. The comparative experiments involving different attention mechanisms and ablation studies revealed that embedding the CBAM into the generator effectively increased the network's focus on critical disease-related features, resulting in more realistic and clearly defined disease-affected regions in the generated images. Furthermore, the introduction of the perception similarity loss function substantially enhanced the network's ability to perceive and represent disease-related information, thereby improving the visual fidelity and realism of the generated images. Additionally, transfer learning applied to object detection models such as YOLOv5s, YOLOv7-tiny, and YOLOv8s led to significant improvements in disease detection performance on the augmented dataset. Notably, the detection accuracy of the YOLOv5s model increased from 79.7% to 93.8%, representing a considerable enhancement in both generalization ability and robustness. This improvement also effectively reduced the rates of false positives and false negatives, resulting in more stable and reliable performance in rice disease detection tasks. Conclusions The rice leaf disease image generation method based on the improved CycleGAN model, as proposed in this study, effectively transforms images of healthy leaves into those depicting disease symptoms. By addressing the challenge of insufficient disease samples, this method significantly improves the disease recognition capabilities of object detection models. Therefore, it holds considerable application potential in the domain of leaf disease image augmentation and offers a promising new direction for expanding datasets of disease images for other crops.

Key words: rice leaf disease, data enhancement, CycleGAN, CBAM, perceptual similarity loss, transfer learning