Robust Peanut Size Grading via Scale-Augmented Ordinal Classification and Morphological Feature Fusion

Deep learning models classify peanut size accurately under fixed imaging conditions, but accuracy drops when apparent object scale changes, for example with camera-to-object distance. Ordinal losses and feature-level fusion are rarely compared under a leakage-controlled protocol. To address both gaps, this study evaluates a single pipeline that pairs scale-aware training with morphological feature fusion. Five backbone architectures (ViT-Base, DeiT-Small, Swin-Tiny, ResNet-50, EfficientNet-B0) were trained with three loss functions (cross-entropy, CORAL, CORN). Continuous isotropic scale augmentation (50–150%) simulated the scale change expected in the field, and robustness was measured with a ten-step synthetic scale scan against a non-augmented control. Independently trained CORAL and CORN outputs were then stacked with analytically rescaled morphological features, to test whether fusion adds genuine value beyond the base models. In the resulting comparison, transformer backbones achieved higher classification accuracy than convolutional networks, and CORN was the most stable of the three loss functions. Scale augmentation delivered a clear robustness gain at a modest cost: for ViT-Base with CORN, mean accuracy across the tested scale range rose from 43.3% to 88.5%, and cross-scale variability fell sevenfold, with only a small loss at native scale. The same pattern held for Swin-Tiny (42.9% to 89.2%, a 7.6-fold reduction) and ResNet-50 (43.4% to 80.3%, a 5.4-fold reduction). Gradient-boosting fusion then raised accuracy by 7.0 percentage points on average over the best single model, with the largest gains on the weaker convolutional backbones. Permutation importance and boundary-case analysis confirmed that this gain did not rest on a single dominant size-correlated feature. Scale-aware training and analytically consistent fusion, rather than architectural complexity, therefore account for the main improvement in image-based ordinal size grading.