High-Definition Light and Shadow Reconstruction of Blast Furnace Burden Surface Images Based on a Spatiotemporal Multi-Domain Guided Generative Model

The blast furnace (BF) is the most energy-consuming and emission-intensive equipment in the steel industry. Burden surface images provide reliable and concrete information that is crucial for real-time monitoring, understanding burden distribution, and ensuring efficient and green operations. However, the high-temperature, enclosed, dusty, and low-light environment within the BF results in images captured by high-temperature industrial endoscopes that suffer from unevenly illumination, abnormal light and shadow, and blurriness, limiting their application. This paper proposes a method for high-definition reconstruction of BF surface images based on a spatiotemporal multi-domain guided generative model (STMGG). This method constructs a deep image prior (DIP) network with an encoder-decoder mechanism to decompose the original burden surface image into illumination and reflectance maps across multiple domains. For the illumination maps, a multi-frequency-attention fusion block (MFAFB) extracts extreme bright and dark features from the original image, calculates the appropriate fusion weights for the illumination maps of each domain, and synthesizes an optimal distributed total illumination map to accurately restore light and shadow. For the reflectance maps, global temporal and multi-spatial domain local structural texture information guidance based on wavelet transform is introduced, along with a structural texture entropy loss, to ensure the generated reflectance maps focus on both global contours and local details, thus reducing blur. The final high-brightness, high-definition reconstruction of the BF burden surface is achieved by merging the total illumination and reflectance maps. Experimental results demonstrate that STMGG significantly enhances the quality of BF burden surface images and exhibits strong adaptability in general scene images, highlighting its broad applicability and significant practical value.