With the rapid development of image editing and generation technologies, forgery techniques have become increasingly sophisticated and covert, posing severe challenges to image tampering localization. Existing methods still exhibit deficiencies in feature representation, feedback mechanisms and adaptability. To address these issues, this paper proposes a novel image tampering localization network, termed PFMF-Net, which integrates coarse localization, progressive feedback, and multi-feature fusion into a unified framework. The network consists of three branches: a coarse-grained localization branch that quickly obtains the initial distribution of forged regions to provide prior guidance for subsequent processing; a progressive feedback optimization branch that guides gradual refinement of features to alleviate semantic degradation; and a multi-feature fusion branch that aggregates complementary features via dual-axis attention to enhance boundary and detail awareness. Experimental results demonstrate that PFMF-Net outperforms existing methods on multiple benchmark datasets, particularly exhibiting higher robustness and accuracy in complex tampering and cross-platform scenarios, thereby validating the effectiveness of the multi-feature fusion strategy and progressive feedback optimization mechanism in improving tampering localization accuracy.