ProtoRefine: Prototype-conditioned refinement of strawberry instance masks with native-area decoding

Abstract In strawberry canopy images, leaves and stems can obscure fruit contours, leaving correctly localized instances with masks that include vegetation or miss visible fruit regions. We propose ProtoRefine to correct these post-detection errors by reusing the instance representations of a prototype-based segmenter, without an additional image encoder. A shared local predictor combines spatial prototypes, instance coefficients, coarse mask logits and coordinates to predict foreground scores. Native-area decoding then preserves the native mask's foreground count and reallocates pixel locations according to these scores, while leaving boxes, classes and confidence scores unchanged. On the official StrawDI test set at 640-pixel input, ProtoRefine achieves 84.03 mask average precision (AP), improving native YOLO26s-seg by 1.32 percentage points in AP and 2.90 points in AP90, with only 25,144 additional trainable parameters. Controlled comparisons across three local-head initializations show that the additional AP benefit of explicit prototype factors over coarse-field inputs depends on decoding and is larger under free thresholding. Matched-instance analyses show higher mean intersection over union (IoU) and boundary F1 than native masks at unchanged foreground count, while an inaccurate native count limits correction. These results show that existing instance representations can support compact local mask correction, and that the incremental benefit of explicit prototype inputs depends on mask reconstruction.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11441508/v1
Primary Topic
Smart Agriculture and AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

ProtoRefine: Prototype-conditioned refinement of strawberry instance masks with native-area decoding

Jiayun Zheng, Minhui Kang, Yiyao Zhang, Peizhe Long et al.
Research Square
Smart Agriculture and AI
preprint

ProtoRefine: Prototype-conditioned refinement of strawberry instance masks with native-area decoding

Jiayun Zheng, Minhui Kang, Yiyao Zhang, Peizhe Long, Jiafeng Qi
preprint en

Abstract

Abstract In strawberry canopy images, leaves and stems can obscure fruit contours, leaving correctly localized instances with masks that include vegetation or miss visible fruit regions. We propose ProtoRefine to correct these post-detection errors by reusing the instance representations of a prototype-based segmenter, without an additional image encoder. A shared local predictor combines spatial prototypes, instance coefficients, coarse mask logits and coordinates to predict foreground scores. Native-area decoding then preserves the native mask's foreground count and reallocates pixel locations according to these scores, while leaving boxes, classes and confidence scores unchanged. On the official StrawDI test set at 640-pixel input, ProtoRefine achieves 84.03 mask average precision (AP), improving native YOLO26s-seg by 1.32 percentage points in AP and 2.90 points in AP90, with only 25,144 additional trainable parameters. Controlled comparisons across three local-head initializations show that the additional AP benefit of explicit prototype factors over coarse-field inputs depends on decoding and is larger under free thresholding. Matched-instance analyses show higher mean intersection over union (IoU) and boundary F1 than native masks at unchanged foreground count, while an inaccurate native count limits correction. These results show that existing instance representations can support compact local mask correction, and that the incremental benefit of explicit prototype inputs depends on mask reconstruction.

Research Square
Jiangxi Electric Vocational and Technical College (CN), Shandong Agricultural University (CN)
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