<p>Accurate recognition of underwater acoustic targets remains a challenging task due to strong background noise, environmental variability, and bandwidth limitations. This study proposes a hybrid framework that integrates a deep wavelet autoencoder (DWAE) with a dilated MobileNet (DiMobNet) to enhance classification reliability under such adverse conditions. The DWAE employs wavelet decomposition to represent acoustic signals across multiple scales, facilitating both noise suppression and the extraction of discriminative time–frequency features. These representations are subsequently processed by a lightweight convolutional network with dilated kernels, which expands the receptive field and improves contextual awareness without increasing computational complexity. Experimental results on the publicly available ShipsEar dataset demonstrate that the proposed method achieves a recognition accuracy of 98.24%, outperforming conventional CNN-based approaches. Ablation studies further verify the complementary roles of DWAE in denoising and DiMobNet in robust classification, particularly under low signal-to-noise ratio conditions (e.g., 5 dB). Moreover, the proposed model reduces inference time by approximately 22% compared with ResNet-50, highlighting its suitability for real-time applications. Overall, this work effectively combines advanced signal processing techniques with efficient deep learning architectures, providing a scalable and practical solution for robust underwater acoustic target recognition.</p>