Image clearness processing for image restoration based on generative adversarial networks

•Innovative technique combining GANs and multi-scale attention mechanism.•Superior image restoration results and high PSNR values.•Enhanced stability and low error for improved image recognition accuracy. Image restoration and enhancement techniques have found widespread applications in the field of...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:International journal of cognitive computing in engineering Ročník 6; s. 360 - 369
Hlavní autoři: Wang, Xin, E, Na, Yang, Jingbo
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 01.12.2025
KeAi Communications Co., Ltd
Témata:
Enhancement
Generative adversarial networks
Image restoration
Multi-scale attention mechanism
Style rendering
Generative adversarial networks
Image restoration
Multi-scale attention mechanism
Enhancement
Style rendering
ISSN:2666-3074, 2666-3074
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:•Innovative technique combining GANs and multi-scale attention mechanism.•Superior image restoration results and high PSNR values.•Enhanced stability and low error for improved image recognition accuracy. Image restoration and enhancement techniques have found widespread applications in the field of digital image processing, spanning areas such as medical imaging, remote sensing image analysis, and historical document restoration. However, with the advancements in generative adversarial networks, existing technologies are no longer sufficient to meet the high-definition requirements of image restoration. In light of this, the experiment proposes an image restoration and enhancement technique based on generative adversarial networks and style perception. The proposed model synergistically combines the image feature capturing ability of generative adversarial networks with the feature fusion capability of a multi-scale attention mechanism. This aims to address the complexities and information loss encountered in traditional image restoration processes. The results indicated that the proposed method, when applied to the Helen Face dataset, consistently increased the average Peak Signal-to-Noise Ratio (PSNR) values across four models. At a system runtime of 0.387 s, the PSNR value for the proposed method reached 52.84 dB, while the PSNR values for the other algorithms continue to increase. On the CelebA dataset, when the proposed method achieved the maximum Structural Similarity Index (SSIM) value, the corresponding number of restored images was 360, with a continuously rising SSIM value reaching 0.968. A comparative analysis of the four methods for image restoration revealed that the proposed method exhibited the highest degree of consistency with real images, demonstrating superior performance in handling details that closely resemble those in real images. The restoration effectiveness of the proposed method surpassed the other three methods significantly. These results indicate that the proposed method yields the best results for image restoration. Furthermore, the stability and low error of the system operation greatly enhance image recognition accuracy, effectively resolving issues related to identity coherence during the image restoration process.
ISSN:2666-3074
2666-3074
DOI:10.1016/j.ijcce.2025.01.005