Unsupervised Domain Adaptation Image Segmentation for Froth Flotation Process Based on Appearance Generative - Adversarial Adaptation Networks

Deep learning-based approaches have made significant advancements in flotation froth image segmentation. However, these methods often rely on large labeled datasets, which are both time-consuming and costly to annotate. A viable solution involves training models on existing labeled datasets (referred to as the source domain) and applying them to new datasets (the target domain). Nonetheless, simplistic models may struggle with variations across different domains, a challenge known as unsupervised domain adaptation (UDA). To tackle this, we introduce Appearance Generative - Adversarial Adaptation Networks (AG-AAN). In this paper, we employ Appearance Generative Networks to adjust source domain images, generating adapted images that exhibit the style of the target domain while preserving their original semantic information. Adversarial Adaptation Networks utilize adversarial learning to optimize feature extraction and segmentation networks, thereby reducing representation disparities between the source and target domains. We conducted extensive experiments on two datasets collected from actual flotation industry sites: antimony ore images as the source domain and copper ore images as the target domain. The experimental results demonstrate that our methods have considerable performance and broad application prospects.