Improved DCGAN Based on Coordinate Attention for Efficient Image Generation Under Small Sample Conditions

In deep learning, Generative Adversarial Networks (GANs) are highly regarded for their exceptional performance in data generation tasks. However, training GANs typically relies on large-scale datasets and extensive iterations, which poses significant challenges in environments with limited data or computational resources. To address these challenges, this paper proposes an enhanced model, CA-DCGAN, based on Coordinate Attention (CA). The CA mechanism effectively enhances the generator's and discriminator's ability to capture and represent spatial features, enabling the generator to produce more detailed and realistic images. We conducted experiments on a censored MNIST dataset to evaluate the model's generative capabilities in small-sample scenarios. By reducing the number of training rounds and applying a consistent training strategy, we systematically compared the performance of CA-DCGAN with other generative models. The experimental results demonstrate that CA-DCGAN not only captures spatial features more efficiently and generates images with greater fidelity and detail but also achieves faster convergence. Furthermore, the study highlights that even under limited data or computational constraints, CA-DCGAN exhibits significant advantages, offering new theoretical insights and practical solutions for efficient image generation.