The invention relates to the technical field of agricultural intelligent
information acquisition and
crop phenotype analysis, in particular to a wild rice bacterial leaf
blight resistance evaluation method based on RGB-D segmentation and three-dimensional
distance measurement, and the technical scheme comprises the steps of
image acquisition and alignment, scab segmentation,
minimum bounding rectangle and longest side extraction, three-dimensional
length measurement and calculation, resistance grade evaluation, and result filing and query. Wherein the aligned RGB frames are extracted and input into the lightweight segmentation network, the segmentation network comprises a U-Net backbone, a multi-dimensional space attention module and a multi-scale
feature fusion module, the low-resource deployable lightweight segmentation model takes the U-Net as the backbone, after MDSAM and MSHFM are introduced, the segmentation IoU is approximately equal to 94.14%, the recall is approximately equal to 97.20%, advantages are provided for the IoU and the recall in multi-model comparison, Params is approximately equal to 0.85 M,
FLOPs is approximately equal to 4.33G, and the segmentation network has the advantages of being high in robustness, high in robustness and high in robustness. Compared with PSPNet, DeepLabV3 and the like, the accuracy and robustness are improved, light weight and high efficiency are achieved, and real-time or quasi-real-time reasoning of an adaptive edge end and a
mobile end is achieved.