Post-harvest quality grading of betel leaves (Piper betle L.) is critical to supply chain efficiency across South and Southeast Asia, yet assessment remains an entirely manual process susceptible to inter-rater inconsistency and throughput bottlenecks. Visual freshness cues—saturated green pigmentation, surface desiccation, and brown discoloration—form a continuous rather than discrete progression, making threshold-based color analysis insufficient. We propose BeLFG-Net, a Betel Leaf Freshness Grading Network, that addresses this challenge through a hybrid four-block encoder combining standard convolutions for initial feature establishment with depthwise separable pairs for parameter-efficient cross-channel mixing, followed by global average pooling and a compact classification head. The key efficiency gain arises from the depthwise separable factorization, which decouples spatial filtering from channel-wise integration and reduces computation approximately 8–9 × relative to equivalent standard convolutions, yielding a total of only 979,651 parameters. We further introduce a progressive dropout schedule (p ∈ {0.25, 0.25, 0.25, 0.50}) that imposes strongest regularization at the bottleneck transition, effectively curbing overfitting on the 766-image field dataset augmented to 6262 samples. On the three-class freshness discrimination task—fresh, moderate, and stale—BeLFG-Net achieves 95.72% test accuracy with per-class F1-scores of 0.96, 0.96, and 0.95, outperforming ResNet-50 (93.77%) and MobileNetV2 (91.65%) while using 26 × and 3.5 × fewer parameters, respectively. Ablation experiments confirm that the hybrid depthwise separable block design, progressive dropout, and global average pooling each contribute independently to generalization, with combined improvements of