Fine coal flotation is one of the most widely used separation techniques in coal processing and utilization, and the dynamic characteristics of flotation froth serve as key indicators for characterizing variations in flotation operating conditions. To address the challenges of feature extraction from fine coal flotation froth images under complex conditions, this paper proposes, for the first time, a feature encoding network that integrates a feature pyramid with shallow feature enhancement and incorporates it into a keypoint detection framework. Through multi-scale feature fusion and low-level feature enhancement, the proposed network improves the representation capability for froth structures at different scales, significantly enhancing the detectability and stability of feature points in froth images. In addition, a Spatially Adaptive Keypoint Regularization method is designed to address the issue of large local deviations in existing velocity estimation approaches. Furthermore, by integrating the proposed method with LightGlue, high-precision matching and dynamic feature extraction in complex froth scenarios are achieved. Experimental results demonstrate that, on the HPatches dataset, the proposed method achieves a repeatability of 0.605 and a homography estimation accuracy of 0.433. On a self-constructed coal slurry froth image dataset, the repeatability is improved to 58.4%, while the velocity entropy is reduced to 3.441. Moreover, the method exhibits low inference latency and low GPU memory consumption, meeting the requirements for industrial real-time monitoring, and providing more reliable technical support for real-time perception of flotation conditions and performance evaluation of the separation process.