Accurate segmentation of froth images is crucial for online monitoring of flotation performance. However, uneven illumination, specular highlight overlap and low-contrast edges frequently cause over-segmentation and under-segmentation when classical watershed transform is applied. In this study, an improved watershed algorithm incorporates specular highlight overlap correction and multi-directional edge constraints. First, Intuitionistic Fuzzy C-means (IFCM), adaptive thresholding and open-close morphological reconstruction are combined to extract and classify highlights into small, medium and large bubbles. An overlap-correction fusion strategy is then designed to remove spurious markers. Second, horizontal–vertical edges extracted by Laplacian of Gaussian (LoG) and ±45° edges obtained by diagonal gradient filters are merged to generate a complete constraint line. Finally, the fused markers and the edge-preserving constraint line are fed into a marker-controlled watershed, achieving robust segmentation of multi-scale froth images. Experiments on 50 on-site images from a tungsten–molybdenum flotation plant show that the proposed method yields mean Adjusted Rand Index (ARI) values of 81.85 %, 87.63 % and 91.63 % for small, medium and large bubbles, respectively, outperforming two state-of-the-art watershed variants.