A Comparative Evaluation of Selective FP32 and INT8 Poultry Fecal-image Classifiers under Classwise Risk Screening
Year: 2026
Type: article
Abstract: A selective classifier can assign a label or defer an uncertain poultry fecal image.We evaluated MobileNetV2 and EfficientNetB0 using 7,695 images retained from an 8,067-image archive after duplicate and labelconflict screening.Both backbones were trained with three seeds and exported in FP32 and INT8.Validation selected 24 abstention rules; a separate calibration partition screened their classwise automatic-error risk against a ten percent target.Seven MobileNetV2 rules passed; 17 failed the Newcastle bound and were assigned all-defer.On 1,154 heldout test images, the passing rules labelled 68.46 to 76.52 percent automatically.Class-specific thresholds did not improve coverage in either matched setting where both rule types passed.INT8 reduced model size but did not consistently improve host CPU latency.The small Newcastle calibration stratum limited certification despite strong conventional classification scores.Post-hoc sensitivity analyses indicated that Newcastle calibration size and multiplicity accounting, rather than raw accuracy, largely determined certification.
Authors: Nafiatun Sholihah(Conceptualization, Formal analysis, Funding acquisition, Project administration, Supervision, Writing – original draft, Writing – review & editing), Majid Rahardi(Investigation, Methodology, Writing – original draft, Writing – review & editing), Sundari(Writing – review & editing), Ferian Fauzi Abdulloh(Investigation, Software, Validation, Visualization, Writing – original draft, Writing – review & editing)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
SDG: Zero hunger
Open Access status: hybrid
OpenAlex ID: W7235257084
