Semantic‐aware visual consistency network for fused image harmonisation

With a focus on integrated sensing, communication, and computation (ISCC) systems, multiple sensor devices collect information of different objects upload it to data processing servers for fusion. Appearance gaps in composite images caused by distinct capture conditions can degrade the visual quality affect accuracy other image analysis results. The authors propose fused-image harmonisation method that aims eliminate appearance among objects. First, modify lightweight backbone combined with pretrained segmentation model, which extracted semantic features were fed both encoder decoder. Then implement semantic-related background-to-foreground style transfer leveraging spatial separation adaptive instance normalisation (SAIN). To better preserve input information, design simple effective semantic-aware denormalisation (SADE) module. Experimental results demonstrate authors’ proposed achieves competitive performance iHarmony4 dataset benefits from fused incompatible gaps.