In focus stacking and macro photography, gathering maximum information from images captured at different focal lengths without losing realism remains a challenge. This work introduces a novel model called Probabilistic Decision Mask Network (PDMNet) to address this issue by separating the segmentation task into two stages. PDMNet utilizes wavelet transform and is trained end-to-end with a fully focused image to predict a gray-scale decision mask, which represents varying levels of information from different sources. The predicted decision mask isn’t binary; instead, it contains gray-scale values, indicating the amount of information from different sources. The second stage introduces an iterative algorithm to accurately threshold the mask by identifying the highest edge information \(Q_{{\text {AB}}/F}\) metric value, ensuring maximum retention of edge information in the fused image. Experimental results on the dataset multi-focus image fusion in the wild (MFFW) indicate that the method surpasses state-of-the-art, with \(Q_{{\text {AB}}/F}\) of 0.6899, \(Q_{\text {NCIE}}\) correlation at 0.8386, and human visual system \(Q_{\text {CB}}\) at 0.7577.