<![CDATA[Accurate plant disease diagnosis is central to precision agriculture, yet real-world performance degrades under blur, low resolution, and domain shift, weakening zero-shot recognition of unseen diseases. This paper investigates the integration of Coordinate Attention (CA) and Zero-Shot Super-Resolution (ZSSR) as test-time plug-ins to a standard Zero-Shot Learning (ZSL) pipeline without using any target labels. Using Plant Village tomato to potato transfer, each target image is super-resolved via a compact, self-supervised SR CNN (50 inner steps with self-ensemble and back-projection) and then standardized to 224×224×3 before feature extraction with MobileNetV2 (global average pooling). A lightweight CA module enhances spatial channel attention, focusing on lesion regions. The visual embeddings (1280-D) are projected into a 300-dimensional, L2-normalized semantic space through a dense, BN, ReLU to dropout head, and class logits are computed as cosine similarity to Word2Vec prototypes. On the target (potato) test set, the proposed ZSL + CA + ZSSR model achieved 86.33% accuracy, outperforming both ZSL + ZSSR (79.04%) and the ZSTL benchmark (78.34%, VGG16 + Triplet + DAC-300). Confusion matrices show fewer PEB↔PLB and PH to diseased confusions, while training curves exhibit faster, more stable convergence when ZSSR and CA are jointly applied. These results indicate that per-image, test-time ZSSR with CA attention sharpens lesion cues and enhances cross-crop transfer, providing a lightweight, label-free pathway to improved field robustness and diagnostics.]]>
Copyrights © 2026
