A froth-flow feature detection method based on an improved ORB in the NSST domain was de. veloped and applied to flotation dosing state recognition to solve the problems of continuous movement, light effects, and noise interference of flotation surface images, which lead to difficulties in flow feature de-tection. First, two adjacent froth images were decomposed through NSST. Multiscale high-frequency sub-bands were denoised using a scale correlation coefficient and then divided into multiple inner and outer lay. ers. The points of interest were subsequently extracted through modulus maxima detection in each inner layer, and the feature points were extracted through non-maximum suppression between the upper and lower layers. Second, a multiscale BRIEF descriptor was adopted to describe these feature points, the search matching area was dynamically adjusted according to the movement trend of the bubbles. The froth-flow features were then calculated based on the matching results. Finally, a line-and-column autoencoder ex. treme learning machine was constructed to fuse the foam shape, size distribution, and flow features, and the dosing state was recognized by the adaptive random forest method. The experimental results showed that the improved ORB was slightly affected by noise and illumination. The flow feature detection efficien. cy and the detection accuracy were significantly better than those of existing methods. The proposed meth. od could characterize the flow characteristics of the froth surface accurately in different dosing states. The average accuracy of dosing state recognition reached 97. 85%, which was significantly higher than those of existing methods. This study lays a foundation for future research on dosing quantity optimization control.