AGANet: Attention-Guided Generative Adversarial Network for Corn Hyperspectral Images Augmentation

Hyperspectral imaging represents a spectral technique that facilitates the non-destructive detection of corn seeds. However, the application of deep learning techniques often necessitates a substantial volume of training data, a requirement that becomes challenging due to the limited availability of labeled hyperspectral images, a limitation imposed by equipment and labor costs. To overcome these challenges, this paper proposes an attention-guided generative adversarial network designed for data augmentation. By integrating an attention module and a classifier within the GAN framework, this approach enables the generator to produce hyperspectral samples with precise class labels from random noise and class labels as inputs. Additionally, the classifier establishes a linkage between the generated images and their corresponding labels, thereby reducing the necessity for extensive labeled datasets. Moreover, the incorporation of attention modules in both the generator and discriminator enhances the spatial feature extraction, resulting in more realistic sample production. The experimental outcomes confirm that the proposed methodology significantly elevates the resemblance of the generated hyperspectral corn images to actual images, thereby underscoring its potent generative capabilities.