A physics-informed eigenfilter for artifact removal in ultrasonic scanning videos for structural inspection

Artifact suppression in industrial ultrasound videos is critical for enhancing the visibility of weak structural features such as flaws. Traditional approaches such as frequency-based filtering and baseline subtraction are often limited by assumptions of perfectly separated signal components or fixed spatiotemporal alignment, which rarely hold in practice. More recent studies for artifact removal in medical ultrasound have considered eigenspace filtering that relies on the different spatiotemporal coherence between the consistent artifacts (clutter in medical imaging) and the transient signals of interest. This paper adapts the eigenfiltering approach to industrial wheel inspection system implementing a Synthetic Aperture Focus Technique (SAFT) applied to flaw imaging in railroad rails using a Rolling Search Unit (RSU). The paper clarifies key aspects of eigenfiltering applied to industrial ultrasound videos. First, it is shown that the eigenfilter effectiveness primarily stems from the large spatiotemporal autocorrelation (rather than cross-correlation) of the pseudo-stationary artifact reflections compared to the small autocorrelation of the transient flaw reflections. In addition, to address the challenges of shifting artifact positions during a scan (a very common occurrence in practical tests), the paper proposes a novel recursive eigenfilter with rectification that is different from traditional eigenfiltering. This recursive algorithm leverages a nonnegativity constraint consistent with the physics of ultrasound imaging, which iteratively reshapes the eigenspace to optimally suppress the spatiotemporally correlated artifact reflections while highlighting the uncorrelated flaw reflections. The algorithm offers excellent convergence. Experimental results obtained from RSU scanning of both artificial and natural rail flaws demonstrate outstanding filtering performance in the presence of strong artifacts. This filtering approach is widely applicable to many imaging applications involving a scanning setup.