Arrow Research search

Author name cluster

Jiwon Ryu

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

2 papers
1 author row

Possible papers

2

EAAI Journal 2026 Journal Article

Enhancement of apple defect identification with semantic segmentation and label-efficient data generation

  • Jiwon Ryu
  • Sang-Yeon Kim
  • Chang-Hyup Lee
  • Gyumin Kim
  • Harin Jang
  • Taehyeong Kim
  • Suk-Ju Hong
  • Geon Hee Kim

The increasing demand for automated fruit-sorting systems has driven the development of machine vision and deep learning technologies for postharvest grading of fruit defects. Effective sorting requires identifying not only the presence of defects but also their types and severity to avoid unnecessary rejection of marketable fruits with minor defects. This study applied deep learning-based semantic segmentation models to Fuji apple images, focusing on four defect types: cracks, bruises, diseases, and scars. Model performance was evaluated for both defect classification and severity estimation. To further improve performance, a label-efficient approach using generative adversarial networks was proposed to generate synthetic apple images and defect masks, reducing the need for extensive manual labeling to create a larger dataset. Qualitative and quantitative analyses of the generated results showed that the synthetic dataset successfully mimicked the biological characteristics of apples as well as the shape, position, and size of the defects. The segmentation model's ability to identify defects was enhanced by the proposed synthetic dataset. The R2 values for defect severity estimation increased to 0. 82, 0. 85, 0. 75, and 0. 92 for cracks, bruises, diseases, and scars respectively, while F1-scores for defect classification reached 100, 94. 3, 94. 1, and 89. 7 %. Furthermore, per-sample classification performance was enhanced with a binary F1-score of 95. 9 % for defect presence and a multi-label accuracy of 93. 9 % for defect types. This study clearly demonstrates that synthetic datasets generated using generative adversarial networks can substantially enhance both defect type classification and severity estimation in semantic segmentation-based apple defect identification models.

Details DOI

NeurIPS Conference 2025 Conference Paper

CodeGEMM: A Codebook-Centric Approach to Efficient GEMM in Quantized LLMs

  • Gunho Park
  • Jeongin Bae
  • Byeongwook Kim
  • Baeseong Park
  • Jiwon Ryu
  • Hoseung Kim
  • Se Jung Kwon
  • Dongsoo Lee

Weight-only quantization is widely used to mitigate the memory-bound nature of LLM inference. Codebook-based methods extend this trend by achieving strong accuracy in the extremely low-bit regime (e. g. , 2-bit). However, current kernels rely on dequantization, which repeatedly fetches centroids and reconstructs weights, incurring substantial latency and cache pressure. We present CodeGEMM, a codebook-centric GEMM kernel that replaces dequantization with precomputed inner products between centroids and activations stored in a lightweight Psumbook. At inference, code indices directly gather these partial sums, eliminating per-element lookups and reducing the on-chip footprint. The kernel supports the systematic exploration of latency–memory–accuracy trade-offs under a unified implementation. On Llama-3 models, CodeGEMM delivers 1. 83x (8B) and 8. 93x (70B) speedups in the 2-bit configuration compared to state-of-the-art codebook-based quantization at comparable accuracy and further improves computing efficiency and memory subsystem utilization.

PDF Details
v2026.09.13