Answers are grounded in this paper and cite their sources.
Textual descriptions generated by vision-language models (VLMs) improve crop disease recognition, but the text pipeline is unavailable at deployment: the application supplies images only. This paper proposes LITE (Learning To Imagine Text Embeddings), which converts VLM-generated text into privileged training-time information. A multimodal teacher model built upon ShuffleNetV2-x2.0, with a frozen BERT text branch and cross-modal attention, is first trained and then frozen. Subsequently, a lightweight image-only student network is distilled to imitate the teacher. Its text simulator reconstructs the teacher’s text features from image features via a deliberately lightweight three-layer network; the generated imagined text embeddings are aligned with the teacher’s ground-truth text features under a compound loss objective. On two public benchmarks, the student model achieves 98.84% accuracy on the Soybean Disease dataset, compared with the teacher’s 98.99%. The student contains only 2.39 million parameters, representing a 64% reduction relative to the teacher, and attains 67.51% accuracy on the PlantDoc dataset. The proposed method attains the best observed mean accuracy in all comparative experiments on both benchmarks, significantly exceeding the weaker baselines while remaining statistically comparable to the strongest ones. Ablation studies demonstrate that removing the text simulator causes a 2.27% accuracy drop on PlantDoc, while substituting the multimodal teacher with an image-only teacher reduces student performance by 3.98%. The deployed student model requires no text input, VLM inference or network connectivity, fully satisfying the image-only inference constraint.
No takes yet. Share an insight, caveat, or question.
Qiu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: