Evaluating unified training optimisations for MobileNetV2: Efficiency-accuracy trade-offs in fine-grained dog breed classification

Abstract

Deploying convolutional neural networks in resource-constrained settings often requires sacrificing accuracy for efficiency, particularly under small-data regimes. MobileNetV2 is a leading lightweight backbone, but its performance on fine-grained classification tasks can suffer when training data is scarce and compute budgets are limited. This study proposes a unified training optimisation pipeline for MobileNetV2 that integrates five complementary techniques: transfer learning, dynamic data augmentation, dropout regularisation, adaptive learning-rate scheduling, and early stopping. This approach was evaluated on a subset of the Stanford Dogs dataset (1290 images, eight breeds) and compared against a standard MobileNetV2 baseline as well as MobileNetV3Small and EfficientNetB0 under identical training conditions. Performance is assessed via Top-1 accuracy, training/inference time, and statistical significance (ANOVA, Tukey’s HSD, Cohen’s d), with an ablation study isolating the contribution of each component. The optimised pipeline reduces training time by 50–70% (609 s vs. 1213 s) and inference latency by 10% (0.028 s vs. 0.031 s per image), while incurring only a 1.4% drop in Top-1 accuracy (98.6% vs. 100.0%). Statistical tests confirm significant efficiency gains (p < 0.001) and marginal accuracy loss (p = 0.0008). Ablation analysis demonstrates that no single technique achieves these combined benefits. Under identical conditions, MobileNetV2 outperforms alternative lightweight models that fail to converge on this small dataset. Coordinating established optimisation strategies yields a robust, resource-efficient classifier well suited to edge-computing and on-device vision tasks with limited data. This pipeline offers a practical blueprint for deploying deep learning models in small-data, low-power environments.