Identification of highly similar barley varieties using color priors and adaptive enhancement

Abstract

Malting barley (Hordeum vulgare L.) varietal purity is important for stable beer quality, but major Chinese cultivars often show highly convergent seed phenotypes, with subtle seed-coat color differences. Such subtle variation makes rapid, non-destructive identification difficult for conventional convolutional neural networks. We propose Adaptive Hue-Enhanced EfficientNet (AHE-Net), a dual-stream deep learning architecture that combines color priors with adaptive Hue enhancement. AHE-Net uses unsupervised K-means clustering of normalized Hue values in the hue-saturation-value (HSV) color space to derive color-prior centers and applies a target color stretching module for nonlinear Hue remapping. On a dual-view dataset of 14,810 images from 10 barley varieties, AHE-Net achieved 98.01 ± 0.14% accuracy across six paired runs based on the same set of six random seeds, outperforming the EfficientNet-B4 baseline by 1.01 ± 0.18% points. Five-fold image-level stratified cross-validation further showed consistent improvement, with AHE-Net achieving 97.87 ± 0.25% accuracy compared with 96.29 ± 0.26% for the baseline. These results indicate that color-prior-guided Hue enhancement improves the recognition of highly similar barley varieties and supports low-cost, non-destructive raw-material screening under controlled imaging conditions.