Diagnosis and Mitigation of INT8 Quantization Failure in MobileNetV3 for Plant Leaf Disease Classification

Abstract

Edge deployment of deep learning models for plant disease classification requires model compression to fit resource-constrained hardware, yet post-training quantization (PTQ) may introduce accuracy loss that is not always well characterized for compact architectures. This progress report presents interim results from a wider study optimizing a fine-tuned MobileNetV3-Small model for tomato leaf disease classification. After training reached 99.71% validation accuracy on a 10-class PlantVillage subset, standard INT8 static PTQ using ONNX Runtime and TensorFlow Lite produced a severe accuracy collapse (10.4-27.4%) despite systematically testing eleven quantization configurations, including per-channel/per-tensor granularity, calibration methods, operator exclusion, and export settings. Diagnostic analysis revealed a persistent class-wise failure pattern rather than random noise, and an explicit shape-mismatch error confirmed a quantizer defect in per-channel bias requantization for grouped convolutions. As a mitigation, dynamic weight-only and dynamic-range quantization schemes were adopted, restoring laboratory accuracy to 99.4% with 29.3-72.0% size reduction, although field-condition accuracy on secondary open field data dropped to 43.1-43.9%, indicating a substantial domain generalization gap that motivates the next research phase.