DeepWaveFuse: Hybrid Deep Learning and Wavelet-Based Multi-Focus Image Fusion Method

This study presents DeepWaveFuse, a hybrid method that combines deep learning with wavelet techniques to fuse multi-focus images. The method integrates a modified VGG-19 network and attention-based feature extraction with a three-level Meyer wavelet transform to preserve semantic content and high-frequency details. The adaptive fusion method was used to fuse deep semantic features with wavelet domain data, and then guided filtering was used to enhance the fused image and remove fusion artifacts. Four benchmark datasets were used to evaluate the method: Lytro, MFI-WHU, COCO 2017, and PASCAL VOC 2012, including light-field, natural, and synthetically generated multi-focus images. The method achieves peak signal-to-noise ratio (PSNR) scores of 32.5 dB, 33.1 dB, 31.9 dB, and 32.1 dB, with corresponding edge-based fusion metric (QAB/F) values of 0.95, 0.96, 0.94, and 0.95, on Lytro, MFI-WHU, COCO 2017, and PASCAL VOC 2012, respectively. The proposed method takes around 1.2-1.4 seconds to process an image pair using a single GPU. The results show that fusing a multi-focus image using both semantic feature extraction and multi-scale wavelet-domain detail preservation is effective. Ablation studies were further conducted for the attention mechanism, wavelet decomposition, and fusion strategy.