Quantization Trade-offs in Edge AI: Evaluating Lightweight Neural Architectures for Mobile-Based Healthcare and Agricultural Diagnostics

Abstract Edge Artificial Intelligence (Edge AI) enables real-time, privacy-preserving diagnostic decision-making directly on resource-constrained mobile and embedded platforms, eliminating reliance on cloud infrastructure. However, deploying deep neural networks in off-grid agricultural fields and clinical point-of-care (PoC) healthcare environments requires navigating severe trade-offs between predictive accuracy, inference latency, dynamic RAM utilization, and hardware energy constraints. This study presents a rigorous empirical evaluation of model compression techniques across lightweight Convolutional Neural Networks (MobileNetV2, MobileNetV3Small, ShuffleNetV2 x0.5) and nano-scale object detectors adapted for image-level classification (YOLOv8n, YOLOv11n). Using Post-Training Dynamic 8-bit Integer (INT8) Quantization and magnitude-based structured filter pruning, we systematically analyze diagnostic performance across five benchmark datasets spanning mobile healthcare (ISIC 2019 / HAM10000 skin lesion screening, NIH Malaria blood smear cell pathology) and precision agriculture (PlantDoc, Mango, Soybean). Experimental results benchmarked on an ARM-based edge platform (Raspberry Pi 5) demonstrate that dynamic INT8 quantization consistently preserves baseline representational accuracy across both clinical and agricultural domains while reducing storage footprints by over 50%. MobileNetV2 achieves the optimal trade-off between baseline accuracy and quantization resilience, maintaining 90.00% post-INT8 accuracy on complex PlantDoc field data and 92.40% on ISIC 2019 skin lesion screening. Conversely, ShuffleNetV2 x0.5 delivers ultra-low hardware latency (7 ms) and minimal memory overhead (2.4 MB RAM), but suffers severe accuracy degradation (dropping up to 25.15%) under structured channel pruning on visually complex, noisy diagnostic inputs. Based on these empirical findings, we establish a hardware-aware dual-domain architectural selection matrix, a standardized 4-stage edge compression deployment protocol, and actionable solutions for real-world Edge AI integration.