A team led by Associate Researcher CHEN Lei from Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences, has recently proposed an innovative domain adaptive method for rice disease image recognition.
This research achievement was published in Computers and Electronics in Agriculture.
In recent years, transfer learning has been widely studied for crop disease image recognition tasks with limited samples. However, achieving satisfactory results becomes challenging when there are significant differences in data distribution between the target and source domains. To address this issue, the research team presented a domain adaptive image recognition method based on an attention mechanism, using rice disease image recognition with small samples as an example.
The method first optimized the weight distribution of rice disease image features within a neural network, enabling the attention mechanism to focus more on features closely related to diseases. Then, by integrating this attention mechanism with a domain adaptation network, the method reduced the differences in feature distribution between datasets, resulting in improved image recognition accuracy. Comparative experiments were designed to validate the effectiveness of the proposed method.
The research results demonstrated that the method effectively addresses the issue of low image classification accuracy caused by significant differences in data distribution between the target and source domains. Remarkably, even as the size of the source domain dataset gradually decreased, the proposed method maintained a consistently high and stable accuracy level. The image recognition accuracies achieved were 95.25%, 91.50%, and 91.25% using three commonly used domain adaptation models, respectively. This research breakthrough offers a new approach for crop disease image recognition under limited sample conditions.
This work sheds light on the application of fusion attention mechanisms in addressing domain adaptation issues in agricultural image recognition tasks.
Domain adaptation network based on CPAM. (Image by CHEN Lei)
CAM visualization examples under different attention mechanisms. (Image by CHEN Lei)
