Abstract:
Smoke generated during fire and other
critical incidents severely degrades the visual quality of
captured imagery, hindering safety monitoring,
evacuation planning, and firefighting decision-making.
Existing smoke- and haze-removal techniques, ranging
from prior-based dehazing to single-branch generative
models, struggle with thick, structureless smoke and often
leave residual artefacts, colour distortion, or require
large paired real-world datasets. Across the reviewed
literature, prior-based dehazing such as the Dark
Channel Prior typically reports Peak Signal-to-Noise
Ratio (PSNR) values of only around 13-17 dB under dense
haze or smoke, while GAN-based desmoking methods
raise this to roughly 18-25 dB yet still report Structural
Similarity Index (SSIM) scores below 0.85 and visible
residual smoke once density is high, and multi-model
pipelines such as DHSGAN incur added computational
cost from training two networks separately. This paper
proposes a two-stage Generative Adversarial Network
(GAN) framework for blind-inpainting-based smoke
removal. The framework decouples smoke localization
from image reconstruction through two cooperating sub-
networks: a Mask Prediction Network (MPN) that
identifies smoke-affected regions, and a Robust
Inpainting Network (RIN) that reconstructs the identified
regions using adversarial and VGG16-based perceptual
losses. This method reviews the relevant literature,
presents GAN fundamentals for completeness, detail the
proposed system architecture, network designs, dataset
considerations, design goals, and loss formulation, and
position the framework against existing smoke- and haze-
removal approaches identified through a structured
literature review. The proposed design is intended as the
basis for a subsequent implementation and empirical
evaluation.
