Comparative Evaluation of CNNs and Transformers for Cauliflower Disease Classification Using an Audited and Harmonized Dataset

Reliable evaluation of cauliflower image classification requires attention to dataset overlap, label consistency, and generalization across image sources. This study compared six convolutional neural networks and two Transformers using an audited and harmonized dataset assembled from five public repositories. Before splitting, exact-duplicate screening, perceptual-hash review of near-duplicates, label harmonization, and exclusions reduced 5394 downloaded files to 3805 original images across 14 classes. The fixed stratified split contained 2681 training, 562 validation, and 562 test images. The main benchmark compared an unaugmented baseline with six configurations varying input resolution and augmentation, yielding 162 runs across three training seeds. A separate evaluation compared four models across five source-held-out folds and matched mixed-source controls, with 60 runs per arm. Twins-SVT-Base had the numerically highest mean test accuracy at 384 × 384 with offline and online augmentation (99.23%, SD 0.10 percentage points) and the numerically highest mean test macro F1-score at 300 × 300 with offline augmentation alone (98.98%, SD 0.22 percentage points). At 224 × 224, offline augmentation improved both metrics for Xception and macro F1-score for MobileNetV3-Large relative to the unaugmented baseline after Holm correction. No tested online-augmentation contrast or architecture comparison within the 384-pixel online-augmented configuration remained significant after correction. Mean macro F1-scores ranged from 38.80% to 78.56% on held-out sources and from 87.55% to 99.80% in the matched controls. FP32 inference benchmarks characterized computational trade-offs, while Grad-CAM illustrated attribution patterns for selected predictions. These findings support evaluating accuracy, uncertainty, efficiency, and source generalization together.

Authors

Publication Details

Journal
Electronics
Published
2026-11-16
DOI
https://doi.org/10.3390/electronics15225322
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Comparative Evaluation of CNNs and Transformers for Cauliflower Disease Classification Using an Audited and Harmonized Dataset

Sovila Srun, Chamroeun Khim, Makara Mao, Bunchhun Chhim et al.
4 links
Electronics
Smart Agriculture and AI
article

Comparative Evaluation of CNNs and Transformers for Cauliflower Disease Classification Using an Audited and Harmonized Dataset

Sovila Srun, Chamroeun Khim, Makara Mao, Bunchhun Chhim, Keovichear Ouk, Nova Liev, Min Hong
article en

Abstract

Reliable evaluation of cauliflower image classification requires attention to dataset overlap, label consistency, and generalization across image sources. This study compared six convolutional neural networks and two Transformers using an audited and harmonized dataset assembled from five public repositories. Before splitting, exact-duplicate screening, perceptual-hash review of near-duplicates, label harmonization, and exclusions reduced 5394 downloaded files to 3805 original images across 14 classes. The fixed stratified split contained 2681 training, 562 validation, and 562 test images. The main benchmark compared an unaugmented baseline with six configurations varying input resolution and augmentation, yielding 162 runs across three training seeds. A separate evaluation compared four models across five source-held-out folds and matched mixed-source controls, with 60 runs per arm. Twins-SVT-Base had the numerically highest mean test accuracy at 384 × 384 with offline and online augmentation (99.23%, SD 0.10 percentage points) and the numerically highest mean test macro F1-score at 300 × 300 with offline augmentation alone (98.98%, SD 0.22 percentage points). At 224 × 224, offline augmentation improved both metrics for Xception and macro F1-score for MobileNetV3-Large relative to the unaugmented baseline after Holm correction. No tested online-augmentation contrast or architecture comparison within the 384-pixel online-augmented configuration remained significant after correction. Mean macro F1-scores ranged from 38.80% to 78.56% on held-out sources and from 87.55% to 99.80% in the matched controls. FP32 inference benchmarks characterized computational trade-offs, while Grad-CAM illustrated attribution patterns for selected predictions. These findings support evaluating accuracy, uncertainty, efficiency, and source generalization together.

ElectronicsVol. 15(22)
National Research Foundation of Korea
Zero hunger
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon