Table 9 Advantages of proposed study over SOTA models
Feature/metric | Proposed slender-CNN | Traditional CNNs | Hybrid models |
|---|---|---|---|
Parameter count | 387,340 (extremely low) | High (MobileNetV3: ~ 3 M) | Very high |
Accuracy (corn/rice/wheat) | 98.45%/87.11%/99.96% | Lower for comparable models | High but with increased computational cost |
Generalization across crops | High; no architectural modifications needed | Moderate; often requires re-training | High but computationally expensive |
Multi-scale feature extraction | Enabled using multi-scale filters in a single layer | Nil | Partially enabled through attention mechanisms |
Computational efficiency | High; suitable for resource-constrained devices | Moderate | Low; high computational requirements |
Flexibility | Robust for various diseases and crops | Limited to specific conditions | Robust but dependent on large-scale resources |
Deployment feasibility | Highly deployable on edge devices | Limited to mid-range devices | Requires high-end hardware |
Scalability | Excellent; works effectively across datasets | Limited | High but with higher cost |