Skip to Main Content
Article navigation
Purpose

This study addresses the computational and accuracy limitations of the Segment Anything Model (SAM) by Meta AI in segmenting microscopic pores on irradiated nuclear fuel surfaces. It proposes an optimized framework combining adaptive prompts, boundary enhancement, and lightweight SAM variants to automate defect analysis—critical for evaluating nuclear fuel structural integrity and safety.

Design/methodology/approach

A three-stage approach improved SAM: (1) Sauvola-based adaptive thresholding for optimizing prompt ratios (2:1 positive/negative); (2) Gradient filtering to sharpen pore boundaries; (3) Evaluation of SAM variants (ViT-H(Huge)/L(Large)/B(Base), MobileSAM, EdgeSAM) for efficiency-accuracy trade-offs using F1 (a vital measure of classification problems) and Intersection over Union (IoU) metrics.

Findings

The 2:1 prompt ratio with gradient filtering boosted F1 by 6.3% and IoU by 7.7%. SAM ViT-H achieved peak accuracy (F1:82.84%, IoU:70.98%) but required the highest resources. MobileSAM balanced performance (F1:77.83%, IoU:60.36%) with lower GPU usage (778MB), while EdgeSAM prioritized efficiency (524MB) at reduced accuracy (F1:70.17%, IoU:54.87%).

Originality/value

This work uniquely integrates threshold-based prompts and gradient filtering to enhance SAM for nuclear fuel defect segmentation, enabling zero-shot adaptation to low-contrast microstructures. It provides actionable insights into model selection (accuracy vs efficiency) and advances automated structural integrity analysis, bridging traditional image processing with cutting-edge vision models for nuclear energy applications.

Highlights
  1. Developed a method for generating prompts that leverage the prior knowledge of “fuel pores having deeper grayscale” using Meta AI’s Segment Anything Model.

  2. Applied this method to segment fuel pore images processed with image processing techniques and validated that this automatic segmentation method achieves better accuracy than traditional threshold segmentation methods.

  3. Compared the inference efficiency of various base segmentation models, concluding that segmentation performance is generally proportional to the computing resource usage.

Licensed re-use rights only
You do not currently have access to this content.
Don't already have an account? Register

Purchased this content as a guest? Enter your email address to restore access.

Please enter valid email address.
Email address must be 94 characters or fewer.
Pay-Per-View Access
$41.00
Rental

or Create an Account

Close Modal
Close Modal