Conditional Diffusion for Flotation Froth Image Generation

Froth images provide critical visual information for characterizing flotation operating conditions. However, due to the high cost of on-site acquisition and annotation, available data are often scarce and offer insufficient coverage of operating conditions. At the same time, existing generative methods struggle to achieve both high fidelity and controllability. This study proposes a class-conditional diffusion-based generation (c-DDPM) approach for flotation froth image generation, where prior-knowledge of operating conditions is encoded as class labels and injected into the denoising process. In addition, the U-Net noise-prediction network is enhanced to strengthen conditional information propagation and improve multi-scale texture modeling, enabling targeted synthesis of froth images. Moreover, with the real test set kept unchanged, generated samples at different injection ratios are used for data augmentation of a six-layer CNN, and the performance gains are evaluated via cross-validation and multiple random-seed trials. The results show that, at the optimal augmentation ratio, the overall accuracy for the four-class task increases by 5.28% and the Macro-F1 score improves by 7.26%, with stable gains across runs. The proposed method can provide more sufficient training data for model development under small-sample conditions and support engineering applications in intelligent monitoring of flotation processes.