Open-Set Recognition of Rice Leaf Diseases: CLIP vs. CNN (Code and Results)

This dataset provides the code and pre-computed result files accompanying the paper "Open-Set Recognition of Rice Leaf Diseases: A Comparative Study of Vision-Language and Convolutional Architectures." It reproduces all experiments comparing a fine-tuned CLIP (ViT-B/32) against convolutional baselines (ResNet-50, EfficientNet-B0) on six rice leaf disease classes under a leave-one-class-out (LOCO) protocol. Included are the training/evaluation scripts, result files reporting five OSR metrics (AUROC, AUPR, FPR@TPR95, OSRC@0.5, OSRC@0.1) with multi-seed repeats and paired significance tests, and Figures 1–4.

Image data are not redistributed; they were merged from two public sources (Mendeley Data, DOI 10.17632/fwcj7stb8r.1; and Baidu AI Studio, datasetdetail/254694) into 10,030 images across six classes, split 70/10/20 (seed 42). Experiments are CPU-based and seed-controlled; edit the paths in config.py before running. Code is released under the MIT License.