Abstract: This paper aims to systematically review the development, key techniques, and major challenges of machine learning and deep learning methods for underwater acoustic target recognition, and to clarify current research hotspots and future directions. It first reviews the pipeline of underwater acoustic signal processing, including data preprocessing, feature extraction, and related representation methods, covering physically meaningful features, time-frequency features, auditory perceptual features, and multi-feature fusion strategies. On this basis, traditional machine learning baseline methods based on models such as Support Vector Machines (SVM) and Hidden Markov Models (HMM) are analyzed, and deep learning architectures led by Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), attention mechanisms, and Transformers are summarized. Meanwhile, the application of transfer learning and data augmentation under few-shot conditions is discussed. Existing studies indicate that underwater acoustic target recognition (UATR) has gradually evolved from shallow statistical learning toward a data-driven paradigm centered on deep neural feature representation, with the combination of time-frequency representations and deep networks becoming a commonly adopted technical route. Multi-feature fusion, transfer learning, and data augmentation can effectively alleviate the problems of low signal-to-noise ratio, limited samples, and class imbalance. Meanwhile, model performance is highly sensitive to dataset partition protocols, and recording-level evaluation provides a more realistic measure of generalization. In addition, accuracy alone is insufficient for evaluating imbalanced multi-class UATR tasks. Future research should focus on standardized benchmarks, more comprehensive evaluation metrics, cross-domain generalization, model interpretability, and lightweight deployment for edge platforms, while promoting a deeper integration of physical priors with deep learning models.