Crop Disease Classification on Inadequate Low-Resolution Target Images

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• High-resolution images are crucial for agricultural image classification.
• Insufficient high-resolution data significantly hinders crop disease classification performance.
• Limited availability of high-resolution images poses a challenge in agricultural AI applications.

Purpose of the Study:

• To develop a crop disease classification network using Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) when high-resolution images are scarce.
• To evaluate the effectiveness of ESRGAN in generating super-resolution crop images for disease classification.
• To improve the accuracy of crop disease classification with limited low-resolution data.

Main Methods:

• Utilized ESRGAN to generate super-resolution crop images from available low-resolution images.
• Applied transfer learning techniques to address the limited number of training samples.
• Compared the performance of ESRGAN-generated images against four other super-resolution methods in a crop disease classification task.

Main Results:

• The fine-tuned ESRGAN model successfully recovered realistic crop information from low-resolution images.
• ESRGAN significantly improved the accuracy of crop disease classification compared to other super-resolution methods.
• The generated super-resolution images proved effective for the agricultural image classification task.

Conclusions:

• ESRGAN is a viable solution for crop disease classification when high-resolution data is limited.
• Super-resolution techniques can effectively augment datasets for agricultural AI tasks.
• The proposed method enhances both image quality and classification performance in challenging scenarios.