A segmentation method based on boundary fracture correction for froth scale measurement

During the mineral flotation process, the size characteristics of the surface froth significantly influence the production indicators of the flotation process. However, the information within froth images is intricate, with occurrences of froth stacking and bubble adhesion, resulting in indistinct boundary information. Existing methods struggle to accurately and comprehensively segment bubble boundaries. This paper aims to devise a segmentation method that precisely delineates bubble boundaries, enabling the measurement of bubble size characteristics and an assessment of both average bubble size and bubble count status. A cascaded decoding branch incorporating omni-dimensional convolutional point rendering is designed, leveraging the rendering concept to reanalyze and categorize misclassified pixels. This recalibration process enhances segmentation accuracy by addressing these inaccuracies. Experiments were performed using an on-site froth industrial dataset, demonstrating the substantial advantage of the proposed method in segmenting and measuring flotation froth scenes. Specifically, the segmentation accuracy reached 92.86%. Additionally, the measurement errors for the average bubble size and bubble count were only 10.3% and 8.6%, respectively.