LBP-based multi-scale feature fusion enhanced dehazing networks

• Published in: Multimedia tools and applications Vol. 83; no. 7; pp. 20083 - 20115
• Main Authors: Zhang, Xinyu, Li, Ying, Li, Jinjiang, Hua, Zhen
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.02.2024; Springer Nature B.V
• Subjects: Aerosols; Algorithms; Computer Communication Networks; Computer Science; Computer vision; Data Structures and Information Theory; Error feedback; Fog; Haze; Image enhancement; Image quality; Light; Modules; Multimedia; Multimedia Information Systems; Neural networks; Semantics; Special Purpose and Application-Based Systems; Target detection; Self-attention mechanism; SOS feature enhancement; Single-image dehazing; Ed-local binary pattern; Inception module with dimension reductions; Multi-scale feature fusion
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-15343-8

Image dehazing is a prior process to perform advanced computer vision tasks such as certain target detection, and the degree of haze residue directly determines the performance of these tasks. Most existing dehazing methods follow a physical model of haze formation and obtain clear images indirectly by estimating global atmospheric light and transmission maps, but methods that rely on this model alone are difficult to use in real-world, complex hazy weather environments. In this paper, we propose an LBP-based multiscale feature fusion network for single-image dehazing to generate clear images directly end-to-end, and a multiscale feature fusion module is designed through an error feedback mechanism to alleviate a large amount of missing key feature information caused by the downsampling operation. A feature enhancement module using a Strengthen-Operate-subtract enhancement strategy is introduced into the decoder to improve the quality of the output image by refining the features of the image to be enhanced with the previously estimated image. The Inception module is added to the skip connection to alleviate the problem of excessive semantic gap in feature information at its two ends by increasing and deepening the width and depth of the network, and a self-attention mechanism is introduced to assign higher weights to key features. The above strategies enable the network to recover the haze images not only on the physical model of the input image; but also deep into the feature space to capture the internal correlation between each pixel. In addition, an improved LBP module is used to help the network obtain clearer detailed information and texture images through channel-by-channel matching of LBP images. Our model achieves a PSNR of 35.15 and an SSIM metric of 0.9905 on the SOTS dataset, and also performs well compared to state-of-the-art methods on images with different haze concentrations.