Resource-Efficient Multifocus Image Fusion Using Attention-Guided Networks and Bayesian Hyperparameter Tuning

Abstract

Multifocus image fusion (MFIF) is important for vision systems that require both high efficiency and reliable all-in-focus perception, especially in deployment-oriented and resource-constrained scenarios. However, existing deep learning-based MFIF methods often suffer from excessive complexity and empirical hyperparameter tuning, hindering practical generalizability. We propose a lightweight MFIF framework integrating attention-driven feature modeling with automated Bayesian optimization. Specifically, a compact network using multiscale depthwise separable convolutions (DWConvs) and dual-attention mechanisms enhances focus-discriminative representations under a constrained parameter budget. Instead of resource-intensive decoders, we design a parameter-free decision strategy leveraging deep spatial frequency (SF) and gradient cues for robust focus discrimination. The resulting maps are further refined via morphological operations and guided filtering to ensure edge consistency. Furthermore, a Bayesian optimization pipeline is incorporated to systematically determine optimal network configurations, replacing error-prone manual tuning. Experiments demonstrate a superior tradeoff between fusion quality and computational efficiency, delivering competitive performance with a significantly reduced computational footprint. In addition, CPU-only inference experiments are conducted to evaluate efficiency under resource-constrained settings, showing the potential practicality of the proposed method in deployment-oriented scenarios. The codes are available at https://github.com/514142/MIMF/tree/main