SwinLPB-Net: a local perception enhanced swin transformer with UPerNet and efficient channel attention for multi-focus image fusion

Multi-focus image fusion (MFIF) plays an important role in digital photography by combining different focus regions from several input images of the same scene into a unified image that preserves all the important information. Lately, deep learning-based methods have become popular in this domain due to their high performance. However, many existing methods struggle to achieve a suitable balance between global contextual understanding and preservation of local features. This frequently results in blurred edges, loss of structural consistency and reduced clarity in fused images. To address these challenges, this paper introduces SwinLPB-Net framework, which uses a Swin Transformer backbone built on the UPerNet architecture. This design leverages the strengths of both transformers and convolutional neural networks (CNNs), with transformers capturing global context and CNNs extracting local features. A Local Perception Block (LPB) is added at each stage of Swin Transformer to improve the network's ability to detect small focus variations while Efficient Channel Attention (ECA-Net) is incorporated in both encoder and decoder to improve global feature representation through adaptive cross-channel attention. Extensive experiments on benchmark datasets show that SwinLPB-Net achieves superior performance compared to 14 state-of-the-art image fusion methods in both qualitative and quantitative measures, using metrics such as QMI, QNCIE, QP, QCV and others. The results show that proposed SwinLPB-Net effectively preserves focus regions and produces high quality fused images.