FD-DEIM: efficient and robust detection of small, irregular foliar lesions in the field.

The automated detection of early-stage foliar diseases in open fields is severely constrained by two theoretical bottlenecks: the physical erasure of high-frequency spatial details during downsampling and the topological mismatch between rigid bounding boxes and irregular lesion morphologies. To systematically overcome these structural limitations, we propose FD-DEIM, a lightweight detection framework optimized for fine-grained feature preservation and dynamic boundary adaptation. The architecture introduces four functionally distinct mechanisms: (1) FD-SRFD, an attention-enhanced stem that anchors sub-pixel coordinates via slicing operations in the initial feature extraction stage, preventing the annihilation of microscopic high-frequency details; (2) DRFD, a deep robust downsampler deployed in the network neck, utilizing a parallel multi-path design to protect high-fidelity spatial features from semantic drift during dimension reduction; (3) FD-Block, an ultra-lightweight fusion module that leverages Partial Convolutions to eliminate computational redundancy while efficiently aggregating multi-scale semantics; and (4) D-FINE, a dynamic offset decoder that forces sampling points to actively deform and tightly hug non-convex, organic fungal contours, fundamentally resolving geometric inductive bias. To ensure ecological validity, we also introduce the RTFD dataset, utilizing generative style transfer to simulate extreme meteorological stressors. Extensive evaluations demonstrate that FD-DEIM achieves a superior AP@50 of 0.667. Crucially, by maintaining an unbroken spatial coordinate preservation chain, the model achieves an