The author proposes a SemMIM framework to enhance cross-modal semantic alignment by injecting high-level semantics into local patch encodings and involving text deeply in the MIM process.
