Publication

Abstract : In digital photoelasticity, circular polariscope (CP)–based isochromatics provide retardation information using a single image acquisition. However, for obtaining both fringe order and isoclinic information, phase-shifting techniques are required. This introduces challenges associated with multiple acquisitions and switching between circular and plane polariscope (PP) configurations. Acquiring high-quality CP images poses challenges due to quarter-wave plate errors arising from wavelength mismatches, and the CP configuration requires both waveplates and polarizers. In contrast, plane polariscope (PP) systems are simpler, as they are based solely on polarizer elements. This work explores the role of deep learning in obtaining CP-based isochromatics from fewer plane polariscope–based acquisitions. A PP2CPNet, a transformer-based framework, is proposed for reconstructing CP images from PP inputs with variable acquisition configurations. This framework includes a generic model that supports a variable number of inputs, while dedicated 2-, 3-, and 4-input PP2CPNet models provide superior reconstruction performance. A novel hybrid adaptive weighted-loss objective is implemented to improve optimization stability and performance. Through extensive experimentation across multiple PP acquisition schemes, the configuration providing the most reliable CP reconstruction is identified. All models were trained on datasets containing diverse specimen geometries and validated using out-of-distribution specimens to assess generalization. While the best-performing transformer model with four PP inputs achieved an average PSNR of 39.94 dB and an SSIM of 0.989, the input-specific model with two PP inputs achieved an average PSNR of 33.68 dB and an SSIM of 0.980. The experimental results demonstrate a robust and efficient data-driven approach that improves the usability of photoelastic stress analysis techniques across varied experimental conditions. This enables reliable qualitative stress interpretation from as few as two PP images, which had hitherto been unexplored.