Adaptation regimes of pretrained neural networks in visually proximate class classification: evidence from the PlantVillage benchmark

Abstract

Classifying visually proximate categories remains a difficult problem in image analysis because inter-class differences are often small, whereas intra-class variability is substantial. The aim of the study was to evaluate how the adaptation regime of a pretrained neural network affects performance in a high-complexity recognition task under a fixed architecture and identical experimental conditions. The working hypothesis assumed that partial fine-tuning of the upper layers of a pretrained convolutional neural network would outperform a regime in which the convolutional backbone remains frozen and only the final classification block is trained. The open PlantVillage dataset, containing 54,303 images and 38 classes, was used as a standardized benchmark; its subject domain was treated as a convenient testbed for complex classification of visually similar states. MobileNetV3Small served as the base model. Two adaptation regimes were compared: a frozen convolutional backbone and partial fine-tuning of the upper part of the feature extractor. The main gain was achieved with partial fine-tuning: validation accuracy increased from 0.9707 to 0.9816, while validation loss decreased from 0.0929 to 0.0576. Class-wise analysis on the independent test split showed that the F1-score exceeded 0.95 for 33 of 38 classes, whereas the lowest values, 0.8889 and 0.9078, were observed in groups with high visual similarity. The scientific novelty does not lie in the general idea of partial fine-tuning of a pretrained convolutional network, since this approach is widely used in computer vision, but in the controlled experimental comparison of two MobileNetV3Small adaptation regimes under identical conditions on the PlantVillage dataset. The results show that, for the selected architecture, fixed data split, and visually proximate class classification task, partial fine-tuning of the upper part of the feature extractor improves recognition quality and produces an interpretable error structure.

Plant phenotyping relevance

PlantVillage画像を対象に、植物の視覚的状態を分類するCNNの適応方式を統制条件下で比較・検証しており、画像ベースの植物状態推定手法が研究の中心である。

Code and data availability

The paper uses the public PlantVillage dataset via TensorFlow Datasets and TensorFlow's MobileNetV3Small, but provides no authors' code, trained models, or paper-specific data deposit. PlantVillage is a generic third-party benchmark, not a paper-specific asset, and no availability statement or repository URL for the作者的

No evidence-backed public reproduction asset is currently recorded.