<p>Image manipulation localization is a crucial task in multimedia forensics, with the aim of precisely detecting and locating the forged areas within images. Nevertheless, existing approaches frequently encounter issues such as incomplete detection of forged regions and ambiguous boundary responses when dealing with complex scenarios like blurred edges and minor tampering. To address these issues, we put forward a dual-stream progressive image manipulation localization framework based on cross-fusion to enhance edge perception and localization accuracy. This framework adopts a dual-stream architecture consisting of an RGB stream and an edge stream. The RGB stream captures semantic information, while the edge stream focuses on structural details. The two streams are gradually fused in a progressive decoder to achieve multi-stage fine-grained representations of tampered regions. Furthermore, we design a Cross-Stream Progressive Block Attention module (CSPBA), which encompasses a Cross-Stream Feature Fusion module (CSFF) module and a Progressive Block Attention (PBA) module. The CSFF realizes information guidance and complementarity of the two feature streams during the progressive learning stage, while the PBA models channel and spatial attention separately within local blocks. Experimental results on various standard datasets indicate that the proposed method surpasses existing state of the art methods in image manipulation localization.</p>