Refined Multi-Focus image fusion using Multi-Scale neural network with SpSwin Autoencoder-based matting

This paper proposes a novel method based on a Multi-Scale neural network with a SpatialSwin (SpSwin) Autoencoder-based matting for the multi-focus image fusion (MFIF) task. The proposed strategy introduces several innovations to enhance image quality and address challenges like boundary precision an...

Celý popis

Uložené v:

Podrobná bibliografia
Vydané v:	Expert systems with applications Ročník 276; s. 126980
Hlavní autori:	Jiang, Shengchuan, Yu, Shanchuan
Médium:	Journal Article
Jazyk:	English
Vydavateľské údaje:	Elsevier Ltd 01.06.2025
Predmet:	And Hierarchical Feature Enhancement Block (HFEB) Combined Convolutional Blocks (CCBs) Depthwise Separable Convolution (DSC) Image fusion Pyramid Squeeze Attention (PSA) SpatialSwin (SpSwin) Autoencoder Transpose Convolution (TC) Combined Convolutional Blocks (CCBs) Pyramid Squeeze Attention (PSA) And Hierarchical Feature Enhancement Block (HFEB) Depthwise Separable Convolution (DSC) SpatialSwin (SpSwin) Autoencoder Transpose Convolution (TC) Image fusion
ISSN:	0957-4174
On-line prístup:	Získať plný text
Tagy:	Pridať tag Žiadne tagy, Buďte prvý, kto otaguje tento záznam!

Popis
Shrnutí:	This paper proposes a novel method based on a Multi-Scale neural network with a SpatialSwin (SpSwin) Autoencoder-based matting for the multi-focus image fusion (MFIF) task. The proposed strategy introduces several innovations to enhance image quality and address challenges like boundary precision and texture preservation. The Sum of Edge-Weighted Gaussian-based Modified Laplacian (SEWG-ML) focus measure is developed to refine boundary detection, while a Pyramid Squeeze Attention (PSA) module dynamically adapts spatial and channel-wise attention at multiple scales for precise feature refinement. Trimap generation integrates superpixel segmentation and fuzzy k-means clustering for accurate foreground-background separation. The SpSwin Autoencoder leverages Swin Transformers for hierarchical token reduction and includes a Hierarchical Feature Enhancement Block (HFEB) to reconstruct high-level contextual features, ensuring seamless image fusion with reduced artifacts. These contributions result in a robust, high-performance framework. Experimental evaluations on benchmark datasets demonstrate superior performance, with metrics including SSIM (0.982), PSNR (39.60 dB), and MI (7.968), surpassing existing techniques. The proposed method in this paper is especially good at preserving fine details and suppressing artifacts, indicating its potential applicability in surveillance, medical imaging, and object detection.
ISSN:	0957-4174
DOI:	10.1016/j.eswa.2025.126980