A Source-Aware Multi-Task Framework for Non-Destructive Source Identification and Appearance Quality Assessment of Wheat Kernels

Existing studies on non-destructive visual inspection of wheat kernels mainly focus on single-task variety or defect recognition, limiting their ability to jointly exploit complementary surface information, disentangle source-related and quality-related representations, and mitigate error propagation between prediction tasks. To address these limitations, this study proposes a Source-Aware Multi-Task Grain Network (SAMG-Net) for joint wheat-kernel source recognition and appearance-quality classification. First, a task-guided dual-view multi-scale feature fusion mechanism is developed to capture complementary information from paired UP and DOWN kernel images through cross-view feature interaction, task-specific attention, and adaptive multi-scale aggregation, enabling source recognition and appearance-quality classification to emphasize different morphological, textural, and localized defect characteristics. Next, shared/task-specific representation disentanglement and conditional contrastive learning are integrated to reduce source–quality entanglement by explicitly modeling “same source, different quality” and “same quality, different source” relationships. Furthermore, a confidence-gated soft source-conditioned residual modulation mechanism incorporates calibrated source-probability distributions and prediction confidence into adaptive quality-feature modulation, allowing reliable source information to support appearance-quality classification while limiting the propagation of uncertain source predictions. Experiments on the GrainSet wheat subset demonstrate that SAMG-Net achieves Macro-F1 scores of 0.915 for source recognition and 0.921 for appearance-quality classification, exceeding the strongest baselines by 2.1 and 2.5 percentage points, respectively, with standard deviations of 0.002–0.003 across five independent runs. Additional source–quality combination-holdout, robustness, and view-sensitivity analyses indicate favorable performance under the evaluated internal compositional shifts and controlled visual perturbations, while source-probability calibration reduces the Expected Calibration Error from 0.043 to 0.018. Overall, SAMG-Net provides a dual-view RGB framework for jointly characterizing dataset-defined wheat source categories and visible appearance quality, with potential applications in preliminary grain screening, appearance-based grading, and automated sorting.