Infrared Thermography For Seal Defects Detection On Packaged Products - Unbalanced Machine Learning Classification With Iterative Digital Image Restoration

Aug 6, 2026

Upload by Logic

Non-destructive, online seal defect detection is increasingly used in packaging, notably for food and pharmaceutical products, because it prevents costly leakage-related losses and line disruptions. This work combines machine learning with infrared thermography to overcome constraints caused by small, unbalanced training data and optical imperfections. A proposed classification method reaches over 93% accuracy using two small training sets, with 2.5–10 times fewer negatives, while maintaining low computational cost and avoiding prior statistical defect characterization.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

AccessFree to download

Introduction

Sealing as a quality control step

Existing sealing tests and their limits

Infrared thermography benefits and challenges

Machine learning for small, unbalanced datasets

Optical blurring and resolution limitations

Related work

Infrared thermography for seal defects detection on packaged products: unbalanced machine learning classification with iterative digital image restoration Victor Guillot* *Thimonnier, 11 avenue de la Paix, 69650 Saint Germain au Mont-d’Or, France Received 9th of April, 2022; accepted 10th of May 2023 Abstract Non-destructive and online defect detection on seals is increasingly being deployed in packaging processes, especially for food and pharmaceutical products. It is a key control step in these processes as it curtails the costs of these defects. To address this cause, this paper highlights a combination of two cost-effective methods, namely machine learning algorithms and infrared thermography. Expectations can, however, be restricted when the training data is small, unbalanced, and subject to optical imperfections. This paper proposes a classification method that tackles these limitations. Its accuracy exceeds 93% with two small training sets, including 2.5 to 10 times fewer negatives. Its algorithm has a low computational cost, and does not need any prior statistical studies on defects characterization. Key words: Control, seal, machine learning, small and unbalanced training set, thermography, iterative image restoration 1 Introduction To maintain an optimum throughput, the packaging industry has increasing demands for quality control. One of the most critical stages in the process of such industries is the sealing of materials (Figure 1) . Package integrity is usually checked on small test samples, and offline. A small defect in sealing material can induce significant losses if not detected immediately. Such losses include leakage of raw materials used before and after sealing as well as a perturbation on the production line. A minor leakage forces to remove the faulty package out of line, clean the soiled area, and sometimes cancel the post sealing operations like over-wrapping on several batches. Leaks are not the only defects; particle contamination, wrinkles, bubbles, and over-sealing also affect the strength and tightness of the seal, leading to potential issues during the package’s life after sealing. Figure 1 Horizontal sealing on a Vertical Form-Fill-Seal (VFFS) machine Various non-destructive and online tests have been deployed for sealing. Some of these tests rely on 1D variables control like the voltage and current waveforms, as in high-frequency dielectric sealing [1], or the energy transmitted, as in ultrasonic sealing [2] . These methods are easy and quick, but their monodimensional characteristic cannot provide a real 2D integrity image of the seal. On the other hand, several non-destructive 2D sealing control methods also exist. Infrared (IR) thermography is one of the most attractive methods among them. It showcases various advantages like the absence of electrical and mechanical stress and low cost when used with uncooled microbolometers. It can be applied to a variety of materials, including opaque ones in the visible band. Nevertheless, it has certain limitations like low resolution due to a relatively high wavelength, loss of accuracy on moving objects (Figure 2), and compromise of pixel resolution and thermal sensitivity. Some commercial products offer a sealing control with IR like Qipcam from Qipack, 1420 Brainel’Alleud, Belgium. Their performances and image processing methods are not made available in publications. Some of the recent implementations rely on statistical tests and discontinuity detection [3], whose formulations depend strongly on the human experience. Machine learning provides a solution that can mitigate such pre-requisites. In this paper, various learning models are evaluated with small and unbalanced datasets, i.e., having a small fraction of IR images of bad sealings (called negatives) compared to good ones (called positives). This consideration enables a practical and quick implementation of the control whereas the large and balanced datasets r

FAQs

Why is seal defect detection important in packaging processes?

Seal defects can cause raw material leakage and downstream losses, require removing faulty packages, and disrupt production operations and sealing steps.

What limitations make infrared thermography-based classification difficult?

Performance is restricted by small and unbalanced training data and by optical imperfections such as motion blur and optical blurring that reduce effective spatial resolution.

What does the proposed method achieve compared with traditional approaches?

It performs binary classification using machine learning combined with iterative digital image restoration, reaching over 93% accuracy on small training sets with significantly fewer negative samples and low computational cost.

Aug 6, 2026

Upload by Logic