基于图文模型和对比学习的水稻病害识别方法OA

To address the issues of low recognition accuracy and over-reliance on labeled data in traditional rice disease recognition methods based solely on rice images,this article proposes a rice disease recognition method based on an image-text model and contrastive learning.The text modality complements image information,partially mitigating the problem of insufficient image training samples.Contrastive learning,an unsupervised learning approach,improves the current problem of over-reliance on labeled data in training deep learning models for rice leaf disease recognition.Firstly,a rice image-text pair dataset is constructed using the text generation from image technology of large models,GPT-3.5.Then,within the image-text model for rice disease recognition,the text encoder module is optimized to reduce training time.Additionally,a rice disease recognition method based on contrastive learning employs a hard negative sampling strategy within the image encoder of the image-text model.This allows it to effectively learn feature representations of various categories,thereby enhancing model robustness.Experimental results demonstrate that the proposed model outperforms other models in recogni-tion accuracy for rice disease tasks.