Accurate recognition of potato leaf diseases is vital for precision agriculture and food security. Deep learning models have achieved notable progress, yet most rely on the standard cross-entropy loss, which aligns probabilities with labels but overlooks frequency cues that are critical for distinguishing fine-grained disease patterns. Moreover, multi-term objectives often introduce manually tuned weights, increasing sensitivity to hyperparameters and reducing reproducibility. We propose the Frequency-Entropy Coupled Loss (FECL) that integrates cross-entropy with a mutual information term defined between class-specific spectral prototypes and frequency representations of input samples. This unified formulation encourages both semantic correctness and frequency-domain consistency, without requiring manual balancing. From an information-theoretic perspective, FECL functions simultaneously as an entropy regularizer, a kernel alignment mechanism, and a generalization enhancer. Extensive experiments on the PlantVillage potato leaf dataset demonstrate consistent improvements across diverse backbones, including CNNs, VGG19, ResNet18, MobileNetV2, MLPs, and Vision Transformers. FECL enhances accuracy and F1-score compared to cross-entropy, with the most pronounced gains on deeper architectures. Ablation studies confirm the complementary roles of spectral prototypes and mutual information alignment, while robustness tests under blur, brightness shifts, domain shifts, and noise highlight FECL’s ability to generalize in real-world conditions. FECL provides a simple yet effective loss function for plant disease recognition and shows promise for broader applications in domains where spectral consistency is essential.
Full publication URL
