Multi-feature language-image model for fruit quality image classification

Fruit quality classification has a great impact on the modern fruit industry. However, deep learning methods for fruit quality classification often demand a substantial number of labeled samples, which are hard and expensive to collect in many real-world applications, resulting in overfitting and low generalization. The Contrastive Language-Image Pre-Training (CLIP) model, which fuses image and text features, has demonstrated excellent performance in zero-shot classification. Inspired by CLIP, in this paper, we propose a multi-feature language-image (MFLI) model for fruit quality classification, where the fruit image and feature text are fused to enhance feature extraction. Furthermore, we construct a pomelo quality dataset containing first- and secondgrade pomelo. Based on the zero-shot learning results of CLIP on this dataset, we provide recommendations for pre-prompt and multi-feature text. Experimental results show that in both zero-shot, few-shot, and conventional learning sceneries, our MFLI model outperforms state-of-the-art models on seven types of fruits, demonstrating excellent generalization capabilities.