SCB-UNet3+: flotation froth image segmentation via multi-scale context modeling and boundary-aware feature learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

To address segmentation errors caused by variations in bubble size, adhesion between adjacent bubbles, and weak boundaries in flotation froth images, this study proposes SCB-UNet3+,a scale–context–boundary collaborative segmentation network. Built upon U-Net3+, the network integrates Res2Net and coordinate attention in the encoder to enhance multi-scale feature extraction, employs an asymmetric receptive field module at the bottleneck to aggregate contextual information, and introduces asymmetric convolutional residual blocks in the decoder to strengthen weak boundary responses. Multi-scale predictions are further fused to generate the final segmentation results. Comparative experiments, ablation studies, and five repeated runs with different random seeds were conducted on a self-constructed flotation froth image dataset. The results show that SCB-UNet3+ achieves an mIoU of 0.8999± 0.0027 and an F 1-score of 0.8992± 0.0029, with an absolute mIoU improvement of 5.24 percentage points over the baseline. The boundary F 1-score (BF-score) reaches 0.7434, while the mean absolute error in Sauter mean diameter estimation on the test set is 10.9885 pixels. These results indicate that the proposed method improves region segmentation and boundary recovery in challenging flotation froth images.