HPDC-Net is a compact CNN built from three blocks: DSCB (depth-wise separable convolution + pointwise convolution + batch norm), DAPB (randomly choosing max- vs average-pooling in parallel paths, then 1Γ1 channel reduction), and CARB (squeeze-excitation-style channel attention). Trained 200 epochs with Adam (lr 1e-4) on 224Γ224 inputs, it reports 99.40% accuracy / 99.37% F1 on the imbalanced PlantVillage tomato set (18,160 images, 10 classes) and 99.36% on the balanced 11,000-image Tomato Leaf Disease Detection set; a 3-class potato evaluation reached 99.54% .
On its own terms, the efficiency claim is strong: at 0.52M parameters / 0.06 GFLOPs it dominates heavier baselines like the 6.2M-parameter CornerNet that report ~99.97% accuracy β though note those baselines are more accurate; HPDC-Net wins on the accuracy-per-parameter trade-off, not absolute accuracy .
Strengths. The ablation work is genuine: replacing DAPB with only max-pooling drops accuracy to 98.68%, removing pooling layers inflates parameters to 8.03M, and removing the third DSCB drops accuracy to 96.39% β coherent with the architecture logic . Hyperparameter sweeps and ptflops-verified complexity add rigor.
Weaknesses a skeptical reader should flag. (1) Dataset saturation: PlantVillage is famously easy β CNN reviews document accuracy ceilings near 99% on controlled-condition leaf images, and 99.4% vs 99.97% differences are within what single-run variance and split choice can explain; no confidence intervals or multiple-seed runs are reported . (2) The DAPB randomness claim is under-theorized: stochastic per-forward selection between max/avg pooling is unusual; no analysis of gradient behavior, variance across seeds, or whether a fixed choice would match performance is given. (3) Comparison fairness: baselines come from different papers with different splits/preprocessing; "practical suitability" labels are the authors' subjective categorization. (4) Field generalization is admitted but not tested β the conclusion explicitly notes the laboratory-image limitation and targets 30 FPS CPU in future work . Given documented 68-point accuracy drops under domain shift for comparable models, the "robust" in the title is currently unsubstantiated beyond two tomato datasets plus potato.
The lightweight-plant-disease niche is crowded: DTomatoDNet (0.6M params, 98.95β99.34%), EAMultiRes-DSPP (0.83M), and DCNN+LBP fusion (0.4M, 96%) all target similar budgets . Meanwhile, heavy transfer-learning models like EfficientNet-B5 reach 99.91% on broader 39-class PlantVillage data β so HPDC-Net's niche is efficiency, not accuracy leadership. Datasets are public (Mendeley/Kaggle links provided), aiding reproduction, though no released training code is confirmed in the text.
What would change this assessment: field-image validation (e.g., PlantDoc-style data), multi-seed variance reporting, and a head-to-head retrain of baselines under identical splits. Confidence: high on reported numbers (directly extracted), moderate on the efficiency superiority claim, low on real-world robustness.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.