The timely identification of phytopathological threats is critical for food security, yet deploying high-fidelity detection models on resource-constrained edge devices remains a significant bottleneck in smart agriculture due to the inherent trade-off between computational efficiency and feature representation. This work presents LEAFDet, a lightweight yet robust detection framework tailored for real-time eggplant disease monitoring. To support this investigation, a novel dataset, EDD-4K, comprising 4,593 complex field images, was systematically curated. Specifically, we integrate a Dynamic Convolutional CSP Bottleneck, which drastically reduces static parameters by adaptively recalibrating weights instance-wise, ensuring high model capacity with minimal storage footprint. To further optimize computational flow, an ADown module is employed to accelerate feature downsampling without compromising semantic integrity. Crucially, we innovatively propose the Frequency-aware Large Separable Kernel Attention (FF-LSKA) method, which synergizes local textural nuances with long-range spatial dependencies, significantly enhancing feature discriminability for variable lesion morphologies. Benchmarking experiments demonstrate that LEAFDet achieves an AP 50 of 88.93% and an AP 50-95 of 54.98%, significantly outperforming state-of-the-art architectures. The nano-scale variant sustains an inference speed of 115 FPS and field deployment on the CPU of a commercial mobile platform (Snapdragon 8+ Gen 1) confirms stable operation at approximately 23 FPS without thermal throttling, validating the system’s efficacy for real-time, non-destructive disease monitoring.
Lin et al. (Mon,) studied this question.