Integrating high-frequency detail information for enhanced corn leaf disease recognition: A model utilizing fusion imagery

Abstract

There are various types of corn diseases, many of which affect the leaves. However, the specific details such as shape, size, color, and texture of these diseases in images can present challenges for accurate recognition by deep neural networks (DNNs). Furthermore, images of corn leaf diseases captured in the field often contain noise, which can reduce the robustness and effectiveness of the trained model. Addressing these challenges and acknowledging the limitations of current DNNs models in capturing intricate high-frequency details when identifying corn leaf disease images in complex backgrounds, this study proposes a novel corn leaf disease recognition model that incorporates high-frequency information from images. The proposed model enhances the network's fitting capability by integrating high-frequency detailed features from images into the final three layers of the lightweight MobileNetV3-Large architecture. To effectively represent high-frequency information, a high-frequency feature extraction block (HFFE) is devised, and the adaptive ACON-C activation function is employed to enhance the nonlinear expression capacity of high-frequency details. The end-to-end recognition approach yields a 2.1 % increase in average recognition accuracy compared to the baseline MobileNetV3-Large model, indicating that the inclusion of high-frequency information features enhances model performance. Furthermore, experiments introducing varying levels of noise to the test data illustrate the model's superior anti-interference capabilities and robustness. Consequently, our model exhibits significant value and practical utility for real-world applications.