Neural Network Pruning for Lightweight Metal Corrosion Image Segmentation Models

Neural Network Pruning for Lightweight Metal Corrosion Image Segmentation Models

Metal corrosion detection is essential for ensuring structural safety and minimizing economic losses. While deep learning (DL)-based image segmentation has improved corrosion detection accuracy and efficiency, its high computational demands hinder deployment on resource-constrained edge devices. Thi...

Full description

Saved in:
Bibliographic Details
Main Authors: Vincent F. Yu, Gemilang Santiyuda, Shih-Wei Lin, Udjianna S. Pasaribu, Yuli Sri Afrianti
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Computer vision
deep learning
image segmentation
metal corrosion
pruning
Online Access:https://ieeexplore.ieee.org/document/10969780/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Metal corrosion detection is essential for ensuring structural safety and minimizing economic losses. While deep learning (DL)-based image segmentation has improved corrosion detection accuracy and efficiency, its high computational demands hinder deployment on resource-constrained edge devices. This study investigates lightweight DL models for corrosion segmentation by applying structured pruning to reduce computational costs while maintaining accuracy. We evaluate five segmentation architectures (U-Net, U-Net++, FPN, LinkNet, and MA-Net) and three pruning strategies (linear, automated gradual pruning, and movement pruning) on two corrosion image datasets (NEA and SSCS). Detailed trade-off analysis between model size, computational cost (MAC), and segmentation performance (IoU) reveals that pruning up to 90% sparsity leads to a <inline-formula> <tex-math notation="LaTeX">$\leq 10\%$ </tex-math></inline-formula> IoU drop on SSCS and <inline-formula> <tex-math notation="LaTeX">$\leq 5\%$ </tex-math></inline-formula> on NEA, demonstrating that significant compression is possible with minimal accuracy loss. However, some architectures (e.g., LinkNet) and pruning strategies (e.g., movement pruning) show significant performance deterioration, suggesting that pruning effectiveness varies across models. These findings provide insights into optimizing corrosion segmentation models for efficient deployment on edge devices, balancing accuracy and resource constraints.
ISSN:2169-3536

Similar Items