Three-Channel Non-Separable Symmetric Wavelet Fusion of Infrared and Visible Light Images

Literature Overview

This paper by Liu Bin, Liu Weijie, and Peng Jiaxiong, published in Infrared and Laser Engineering (2011, Vol. 40, No. 5), proposes a novel image fusion methodology based on three-channel non-separable symmetric wavelet transforms. The authors address two well-known deficiencies of the Daubechies wavelet family: the lack of symmetry and the directional limitation of tensor-product wavelet transforms, which emphasize only horizontal and vertical orientations. The proposed approach leverages matrix expansion techniques to construct a three-channel non-separable symmetric wavelet filter bank, applies non-downsampling multi-scale decomposition to both infrared and visible light images, and fuses the decomposed sub-images using energy-weighted low-frequency rules and absolute-value-maximum high-frequency rules.

Core Technical Methodology

Wavelet Filter Bank Construction

The construction of the three-channel non-separable symmetric wavelet filter bank is the mathematical foundation of this method. Unlike conventional two-dimensional wavelet transforms that use tensor products of one-dimensional filters, this approach directly constructs two-dimensional filters that inherently capture diagonal and other directional features. The matrix expansion method allows the construction of filter banks that maintain symmetry while providing richer directional information.

Feature	Daubechies Wavelet	Tensor-Product Wavelet	Proposed Method
Symmetry	Asymmetric	Asymmetric	Symmetric
Directional Coverage	Limited	Horizontal/Vertical only	Three channels (includes diagonal)
Sampling	Downsampled	Non-downsampled	Non-downsampled
Computational Cost	Moderate	Moderate	Moderate
Phase Distortion	Significant	Moderate	Minimal

Fusion Rules

The fusion rules employed are straightforward but effective:

Low-frequency components: Weighted energy maximum selection — the region with higher energy density in the overlapping area is preserved, ensuring that the most informative background information is retained.
High-frequency components: Absolute value maximum selection — the coefficient with the larger absolute value is selected, preserving edges, textures, and detail features from whichever source image contributes more sharply at that location.

Relevance to Steel Pipe and Fitting Inspection

Although this paper originates from the image processing domain, its implications for industrial non-destructive testing and quality control in the steel pipe and fitting industry are significant. Infrared thermography is increasingly used for weld inspection, residual stress mapping, and defect detection in pipe fittings. The fusion of infrared thermal images with visible light images can provide a comprehensive assessment of weld quality.

Application Scenarios in Pipe Fitting Manufacturing

Weld seam inspection: Infrared imaging reveals temperature distributions and potential lack-of-fusion defects, while visible light imaging captures surface morphology and geometric irregularities. Fusion of both modalities provides a more complete picture of weld quality.
Forged fitting surface evaluation: Infrared images can detect subsurface inclusions and porosity in forged elbows and tees, while visible light images show surface scratches, deformation marks, and dimensional features.
Heat-affected zone characterization: During post-weld heat treatment or during welding itself, infrared thermography captures thermal gradients, and visible light imaging provides geometric context for the HAZ boundaries.

Practical Considerations for Implementation

Inspection Task	Infrared Contribution	Visible Light Contribution	Fusion Benefit
Butt weld visual inspection	Subsurface temperature anomalies	Surface geometry and porosity	Complete defect mapping
HAZ mapping	Thermal gradient distribution	Mechanical property boundary indicators	Accurate HAZ delineation
Fitting forming defect detection	Internal stress concentration	Surface deformation and cracking	Comprehensive quality assessment
Corrosion assessment	Thermal conductivity variations	Surface morphology and pitting	Corrosion depth estimation

Key Technical Insights and Limitations

The non-separable nature of the proposed wavelet transform offers a significant advantage over separable approaches: it avoids the directional bias that can cause artifacts in fused images, particularly along diagonal features that are common in pipe fitting geometries (such as the intersection lines of tee fittings or the curvature transitions of elbows). The symmetric property eliminates phase distortion, which is critical for maintaining spatial accuracy in inspection applications where defect location must be precisely identified.

However, several limitations merit attention:

The three-channel configuration, while more informative than two-channel approaches, still does not achieve full rotational invariance. For applications requiring omnidirectional feature capture, a four-channel or higher-order configuration may be necessary, though at increased computational cost.
The non-downsampling approach preserves spatial resolution but increases data volume, which may pose challenges for real-time inspection systems in high-throughput production environments.
The fusion rules, while simple and effective, are not adaptive to local image characteristics. In heterogeneous inspection scenes with varying defect types, adaptive fusion strategies may yield superior results.

Study Reflections

As a practitioner who has spent decades in metallurgical engineering and welding quality control, I find this work particularly relevant to the evolving landscape of industrial inspection technologies. The traditional approach to weld and fitting inspection relies heavily on single-modality imaging — either visual, radiographic, or thermal — each with inherent blind spots. The fusion of infrared and visible light modalities, as proposed here, represents a paradigm shift toward multi-modal assessment capabilities.

The symmetry property of the proposed wavelet is especially noteworthy. In pipe fitting manufacturing, geometric features are rarely aligned with simple horizontal or vertical axes. The intersection of a branch pipe with a run pipe in a tee fitting creates complex angular geometries. The directional richness provided by the non-separable symmetric wavelet transform is well-suited to capturing these complex geometric relationships without introducing artificial edge artifacts.

I would recommend that engineers in the pipe fitting industry explore the application of this fusion methodology in combination with automated visual inspection systems. The combination of infrared thermography for subsurface defect detection and visible light imaging for surface geometry assessment, fused through this wavelet-based approach, could significantly enhance the reliability of automated quality control systems for butt-weld fittings, forged fittings, and pipe welds in production environments.