Autoencoder-Enhanced Balanced GAN Framework for Imbalanced Sonar Image Recognition

Deep learning approaches for sonar image recognition exhibit excellent performance under balanced conditions but suffer significant degradation in real-world underwater environments with class imbalance. An integrated framework combining autoencoder with improved balanced GAN is proposed to address this challenge. To mitigate inherent sonar image characteristics including low contrast and high noise, an embedding-enhanced autoencoder with multi-scale loss functions is designed to expand inter-class distances and strengthen feature representation. The trained autoencoder initializes an improved balanced GAN with gradient penalty for high-quality minority-class synthesis. The classifier directly inherits discriminator parameters, leveraging adversarial-trained feature representations. Experimental validation on two sonar datasets demonstrates substantial performance improvements under imbalanced conditions, particularly in severe imbalance scenarios. This approach enhances model robustness in complex underwater environments, advancing underwater detection through data augmentation and transfer learning. The code is available: https://github.com/cwcaiken/sonar.