With the continuous progress of information technology, Virtual Reality (VR) technology has been more and more widely used in the field of education, and the VR classroom has revolutionized the traditional education model due to its unique interactivity and immersion. However, the complex and changing VR classroom scenes bring challenges to effective scene analysis and optimization. Although existing deep learning methods have made significant progress in image processing, they still face the problems of capturing insufficient detail information and under-utilizing global information when accurately segmenting and classifying VR classroom scenes. To address these problems, this study proposes a series of innovative approaches. The first part investigates VR classroom scene segmentation based on feature enhancement and feature distillation. By designing an attention mechanism with multi-pooling compression incentives and a feature dehazing branch structure with "enhance-refine-subtract" strategies, the network's ability to extract valid information is significantly improved and the interference of invalid information is effectively reduced, which greatly enhances the accuracy of semantic segmentation. The second chapter talks about the optimization of VR classroom scene classification based on multi-scale global information enhancement. By incorporating the Transformer structure, multi-scale information is extracted effectively, global associated information is utilized comprehensively, information processing mechanism in the classification process is optimized, and classification performance is enhanced. Results attained in this study not only improves our understanding of VR classroom scenes, but also provides new insights and technical approaches for the application of deep learning models in processing complicated scenes. Moreover, findings of this paper could portend far-reaching implications in the fields of educational technology and computer vision, and broaden the application range of VR classroom.