The rapid proliferation of online learning platforms has made student engagement a crucial metric for evaluating learning effectiveness. However, existing computer-vision-based methods often inadequately address facial occlusion and pose variations caused by changes in camera angle and viewing distance, resulting in unstable feature representations and limiting recognition performance. Moreover, because partial information loss may occur during camera-based facial image acquisition, in this study, we propose a student engagement recognition framework that integrates conditional random masking (CRM) with ConvNeXt-V2 to mitigate the adverse effects of incomplete or disturbed visual sensing signals. The proposed CRM mechanism leverages facial heatmaps to identify key regions and incorporates a random masking strategy, thereby enhancing the model’s ability to learn critical engagement features while improving robustness to occlusion and pose variations. Experimental results demonstrate that the proposed method achieves accuracies of 98.15 and 87.97% on the SED and SVD-OL datasets, respectively. Compared with existing approaches, the proposed method improves overall recognition performance and demonstrates its effectiveness on datasets containing head-pose variations and partial facial occlusion.