Accurate, rapid, and automatic seafloor sediment classification represents a crucial challenge in marine sediment research. To address this, our study proposes a seafloor sediment classification method integrating convolutional neural networks (CNNs) with small-sample multi-beam backscatter data. We implemented four CNN architectures for classification—LeNet, AlexNet, GoogLeNet, and VGG—all achieving an overall accuracy exceeding 92%. To overcome the scarcity of seafloor sediment acoustic image data, we applied a deep convolutional generative adversarial network (DCGAN) for data augmentation, incorporating a de-normalization and anti-normalization module into the original DCGAN framework. Through comparative analysis of the generated versus original datasets using visual inspection and grayscale co-occurrence matrix methods, we substantially enhanced the similarity between synthetic and authentic images. Subsequent model training using the augmented dataset demonstrated improved classification performance across all architectures: LeNet showed a 1.88% accuracy increase, AlexNet an increase of 1.06%, GoogLeNet an increase of 2.59%, and VGG16 achieved a 2.97% improvement.