Recent advances in deep learning have significantly improved image-based methods for plant disease recognition. However, many of these systems are still developed and tested under laboratory settings, which hinders their direct applicability in real-world agriculture. This review examines 43 scientific studies published between 2020 and 2025 that applied DL methods for plant disease recognition. Unlike generalized surveys that classify articles by tasks (e.g., classification or segmentation), this work adopts an architecture-oriented approach. The methods under consideration include YOLO, Faster R-CNN, UNet, CNN+ViT hybrids, and lightweight models such as MobileNet and EfficientNet. Each architecture is evaluated in terms of structure, speed, accuracy, and performance under conditions close to field environments. One of the key objectives of this work is to identify models that are not only accurate but also practical: capable of operating in real time, on resource-constrained devices, and with images directly obtained from the field. This makes the review valuable for researchers and engineers developing AI systems for agriculture