An object-aware adaptive wheat counting method

Abstract

This invention discloses an adaptive wheat ear counting method based on object perception, relating to the fields of agricultural intelligent detection and computer vision. The method includes: extracting multi-scale visual semantic features of wheat ear images at low, medium, and high levels using a backbone network; generating positive samples of wheat ear regions and negative samples of background regions using an information extraction strategy; fusing positive and negative sample features with wheat ear image features across features using an object perception module, and strengthening wheat ear category separation through contrastive learning; integrating multi-scale semantic cues and contextual relationships through dual attention optimization and multi-scale semantic fusion using a hierarchical object refinement module; generating a wheat ear density map by a prediction module, and constructing an objective function by combining contrastive loss and density map loss to complete model training and wheat ear counting. This invention can improve the recognition and counting accuracy of wheat ears at different scales, adapt to complex field environments, different wheat varieties, and wheat ear counting needs at different growth stages, and is suitable for wheat yield estimation, breeding analysis, and agricultural intelligent monitoring tasks in precision agriculture.

Images