Home // ACHI 2026, The Nineteenth International Conference on Advances in Computer-Human Interactions // View article

Semantic Segmentation of Extremely Small Defects in Sliced Apples

Authors:
Yueying Shi
Oky Prima

Keywords: Machine vision; semantic segmentation; small defect regions; extreme class imbalance.

Abstract:
Semantic segmentation of extremely small defect regions in food inspection remains a challenging task due to severe foreground–background class imbalance and the high cost of missed detections. In the inspection of sliced apples, remaining skin and core fragments often occupy only a few pixels, making them prone to being overlooked despite high overall segmentation accuracy. This study systematically investigates semantic segmentation strategies for detecting extremely small defects in ultraviolet (UV) images of sliced apples. A stepwise experimental framework is employed to isolate and evaluate the effects of loss functions, encoder architectures, and decoder designs under identical training conditions. Quantitative results demonstrate that region-based loss functions, particularly Tversky and Focal Tversky losses, provide superior spatial consistency and recall compared to pixel-wise reweighting approaches. Furthermore, lightweight encoders, such as MobileNetV2, combined with UNet-based decoders, achieve more stable and robust performance in preserving fine-grained defect regions across ifferent random seeds. These findings provide practical design guidelines for recall-oriented semantic segmentation in food inspection tasks, where reliable detection of extremely small defects is critical for quality assurance and safety.

Pages: 17 to 22

Copyright: Copyright (c) IARIA, 2026

Publication date: May 24, 2026

Published in: conference

ISSN: 2308-4138

ISBN: 978-1-68558-383-5

Location: Venice, Italy

Dates: from May 24, 2026 to May 28, 2026