Image manipulation localization is essential for ensuring the authenticity and integrity of visual content in fields such as news media, forensics, and cybersecurity. However, existing methods face significant challenges, including limited localization accuracy and poor generalization when handling complex manipulations and cross-domain scenarios. Furthermore, images are susceptible to quality degradation during transmission and compression, degrading localization performance. Additionally, the imbalance between positive and negative samples hinders effective model optimization using traditional loss functions. Building on this, we propose a Wavelet Attention and Hierarchical Feature Fusion Network (WAHF-Net) to enhance tampered representation. Specifically, we design a Wavelet Attention Feature Fusion Module (WAFFM) to enhance edge-aware features via wavelet-based frequency analysis, and a Global Attention Feature Fusion Module (GAFFM) to improve semantic anomaly detection by aggregating high-level contextual cues. To alleviate sample imbalance, we introduce an Adaptive Weighted Loss Function (AWLF), which dynamically adjusts the weights of tampered and real regions, improving sensitivity to sparse manipulations. Experimental results show that WAHF-Net outperforms state-of-the-art methods on multiple public datasets, especially in enhancing edge details and detecting small tampered regions. AWLF further enhances generalization on imbalanced data, making WAHF-Net a robust solution for real-world manipulation localization.