A Novel Blind Restoration Method for Miner Face Images Based on Improved GFP-GAN Model

Abstract

Miner face images, as important carriers of information transmission, are an important means of digital transformation and intelligent management of mining enterprises. In order to address the issue of complex degradation factors such as noise, blurring, and low resolution, a blind restoration model for miners’ face images was proposed based on improved GFP-GAN, and could make it difficult for blind image restoration to balance fidelity and authenticity. Firstly, the model introduced a UNet++ network to remove complex degradation from miners’ face images using the pre-trained StyleGAN2 network as a priori knowledge. Secondly, in the channel-split spatial feature transform layer, a channel attention mechanism was introduced to better use the prior features in the pre-training network, which could make the final output of the miners’ face images consider both authenticity and fidelity. Compared with other model algorithms, the experimental outcomes clearly indicate that our proposed method surpasses the current leading techniques in LPIPS (0.3827), FID (46.51), NIQE (5.206), and other indicators.

Keywords

• Miner face image
• blind image restoration
• super-resolution
• attention mechanism