EZ-Former: An Attention-Guided Native-Resolution Zoom Transformer for Tea Leaf Condition Classification, Benchmarked Against Fifteen Architectures Under a Leakage-Audited Protocol

Symptoms of tea leaf disorders such as red spider mite stippling are millimeter-scale, whereas standard image classifiers see a leaf photograph at 224 px, where much of this detail falls below one pixel. We propose EZ-Former (Evidence-Zoom Transformer), which reads a leaf twice with one pretrained ViT-S/16: a 224 px glance, then six windows chosen at the peaks of the backbone’s own class-token attention and cut from a 672 px input at native resolution. A zero-initialized gated cross-attention fuses the windows into the glance. EZ-Former and 15 baselines (convolutional networks and columnar, hierarchical, hybrid and efficient transformers) were trained with one shared recipe on a public four-class tea leaf dataset (3154 images), using group-aware, leakage-audited splits and five seeds, and additionally with whole acquisition sessions held out. EZ-Former reached a macro-F1 of 0.927±0.009, compared with 0.914±0.020 for the strongest 224 px baseline (MaxViT-T) and 0.886±0.011 for its own ViT-S/16 backbone; dependence-aware sample-level tests confirmed the advantage over 11 of the 15 baselines on the random splits, but not over the four closest ones, and over all six models compared under session-blocked evaluation (over four of them with the eight acquisition sessions as the unit of analysis). When representative baselines were given a higher input resolution (384–672 px), they matched EZ-Former, most of them at lower cost (CAFormer-S18 at 384 px: 0.931±0.013 at 23 GFLOPs), so the gain is a gain of input resolution, which EZ-Former obtains within a columnar ViT at 59 GFLOPs. On an independent six-class field dataset with natural backgrounds, EZ-Former did not differ significantly from any other architecture (macro-F1 0.52–0.55 for the modern ones). Ablations showed that the zoom and its attention-guided placement matter (0.035 and 0.024 macro-F1), whereas the evidence prototypes and the consistency term can be removed.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-19
DOI
https://doi.org/10.3390/app162211488
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

EZ-Former: An Attention-Guided Native-Resolution Zoom Transformer for Tea Leaf Condition Classification, Benchmarked Against Fifteen Architectures Under a Leakage-Audited Protocol

Mehmet Burukanlı
Applied Sciences
Smart Agriculture and AI
article

EZ-Former: An Attention-Guided Native-Resolution Zoom Transformer for Tea Leaf Condition Classification, Benchmarked Against Fifteen Architectures Under a Leakage-Audited Protocol

Mehmet Burukanlı
article en

Abstract

Symptoms of tea leaf disorders such as red spider mite stippling are millimeter-scale, whereas standard image classifiers see a leaf photograph at 224 px, where much of this detail falls below one pixel. We propose EZ-Former (Evidence-Zoom Transformer), which reads a leaf twice with one pretrained ViT-S/16: a 224 px glance, then six windows chosen at the peaks of the backbone’s own class-token attention and cut from a 672 px input at native resolution. A zero-initialized gated cross-attention fuses the windows into the glance. EZ-Former and 15 baselines (convolutional networks and columnar, hierarchical, hybrid and efficient transformers) were trained with one shared recipe on a public four-class tea leaf dataset (3154 images), using group-aware, leakage-audited splits and five seeds, and additionally with whole acquisition sessions held out. EZ-Former reached a macro-F1 of 0.927±0.009, compared with 0.914±0.020 for the strongest 224 px baseline (MaxViT-T) and 0.886±0.011 for its own ViT-S/16 backbone; dependence-aware sample-level tests confirmed the advantage over 11 of the 15 baselines on the random splits, but not over the four closest ones, and over all six models compared under session-blocked evaluation (over four of them with the eight acquisition sessions as the unit of analysis). When representative baselines were given a higher input resolution (384–672 px), they matched EZ-Former, most of them at lower cost (CAFormer-S18 at 384 px: 0.931±0.013 at 23 GFLOPs), so the gain is a gain of input resolution, which EZ-Former obtains within a columnar ViT at 59 GFLOPs. On an independent six-class field dataset with natural backgrounds, EZ-Former did not differ significantly from any other architecture (macro-F1 0.52–0.55 for the modern ones). Ablations showed that the zoom and its attention-guided placement matter (0.035 and 0.024 macro-F1), whereas the evidence prototypes and the consistency term can be removed.

Applied SciencesVol. 16(22)
Bitlis Eren University (TR)
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