THE LIMITS OF COMPUTER VISION IN FDM/FFF PRINTING QUALITY CONTROL AND ALTERNATIVE DIAGNOSTIC METHODS

Abstract

The paper presents a critical review of the limits of computer vision in quality control of FDM/FFF 3D printing and identifies the defect classes for which standard visible-range imaging is fundamentally insufficient. The relevance of the topic stems from the fact that modern RGB cameras, 2D/3D vision systems, and machine-learning models already perform well in recognizing surface and geometric deviations such as layer shift, under-extrusion, over-extrusion, stringing, externally manifested bed detachment, and contour errors. However, the defects that most strongly determine structural performance, sealing, and long-term reliability are often hidden, volumetric, or material-state dependent. Based on current literature on in-situ monitoring, nondestructive evaluation, and polymer-process characterization, the study demonstrates that standard computer vision cannot directly and reliably detect internal porosity without surface manifestation, enclosed voids, kissing bonds, actual interlayer bond strength, residual stresses before visible deformation, changes in crystallinity, thermal degradation, moisture-related material state, or early nozzle clogging and filament-slip conditions when they have not yet affected the visible surface. A defect observability framework is therefore proposed, distinguishing directly visible, indirectly inferable, and non-observable defects from the standpoint of conventional computer vision. For each defect category beyond the effective range of vision-only inspection, appropriate complementary diagnostic methods are mapped: micro-CT for internal geometry and volumetric porosity; active infrared thermography, ultrasonics, optical coherence tomography, and electrical-resistance monitoring for sub-surface defects and bonding anomalies; force sensing, fibre Bragg grating sensors, digital image correlation, and thermo-mechanical digital twins for residual stress build-up and latent warpage; and vibration, acoustic emission, motor-current, temperature, pressure, and encoder signals for early extruder faults. The paper further argues that interlayer strength, crystallinity development, and polymer degradation must be assessed through mechanical testing, calorimetry, spectroscopy, or rheological monitoring, because these quality indicators are not directly observable in optical images. The main contribution is a practically oriented multimodal quality-control concept in which computer vision remains a highly valuable first-line tool for exposed geometry and process anomalies, but cannot serve as a stand-alone acceptance criterion for safety-critical or function-critical FDM/FFF parts.