<p>Accurate and lightweight recognition of potato leaf diseases is important for early field intervention, yet many high-accuracy plant disease models are designed as image classifiers and do not expose an explicit localization interface. This paper presents CA-NWD-YOLOv5n, a compact object detector for leaf-level potato disease recognition. The framework converts the public PlantVillage potato subset into a YOLO detection benchmark with three categories, namely early blight, healthy leaf, and late blight, and redesigns a YOLOv5n baseline through two complementary mechanisms. First, a coordinate-aware C3 block is inserted into the backbone and neck so that the detector can encode channel saliency together with long-range horizontal and vertical positional cues. Second, the bounding-box regression objective is reformulated as an NWD-CIoU mixed loss, where the normalized Gaussian Wasserstein distance stabilizes localization when leaf boundaries are weak, irregular, or affected by background variation. The resulting model contains only about 1.78 million parameters and is trained at an input resolution of 416 pixels. Experiments on 2152 potato leaf images show that CA-NWD-YOLOv5n achieves 0.995 precision and 0.979 recall on the test split, with an inference time of approximately 1.1 ms per image on an RTX 3090 GPU. Visualization with prediction overlays, confusion matrices, precision-recall curves, and Grad-CAM-style activation maps further indicates that the proposed lightweight detector focuses on disease-relevant leaf regions while preserving deployment efficiency. The study provides a reproducible detection-oriented baseline for potato leaf disease recognition and a practical route for upgrading classification datasets into detector-compatible benchmarks.</p>