Abstract:
In the face of complex flotation site environments and issues such as unclear boundaries caused by the mutual adhesion of flotation froth, existing methods for extracting dynamic features (such as flow velocity and collapse rate) often fail to accurately delineate the dynamic feature sampling regions corresponding to each froth, cannot comprehensively match feature points between adjacent frames, and have difficulty effectively identifying collapse regions. To address these problems, a dynamic feature extraction method for flotation froth based on a centroid-convex hull-adaptive clustering approach is proposed. This method employs an improved Mask2Former, integrated with the multi-scale feature extraction capability of Swin-Transformer, to accurately locate froth centroids and effectively identify collapse regions. An optimal convex hull evaluation function is used to search for the convex hull formed by the centroids of adjacent froth surrounding the target froth, thereby fitting a dynamic feature sampling region close to the actual froth contour. The local feature matching with transformer (LoFTR) algorithm is applied to match feature point pairs between adjacent frames. For all feature point pairs within the dynamic feature sampling region, the main flow velocity of each froth is extracted using the main feature adaptive clustering method based on the OPTICS algorithm. Experimental results show that this method achieves accuracy rates of 88.83% and 97.92% and intersection over union (IoU) rates of 77.90% and 96.52% in ordinary froth centroid location and collapse region identification tasks, respectively. It also achieves a correct feature point pair matching rate of 99.93% with an average exclusion rate of 2.69%. The method effectively delineates feature sampling regions close to the actual froth boundaries under various conditions, enabling the quantitative extraction of each froth's dynamic features.