Tiny Defect-Oriented Single-View CT Reconstruction Based on a Hybrid Framework

Defect detection plays an important role in industry quality control and process evaluation. Learning-based computed tomography (CT) detection with 3-D volume restoration from a single X-ray image is a promising method faster than conventional CT reconstruction. Although the existing methods have shown that 3-D shapes can be well reconstructed from a single X-ray image, the restoration of small internal details, especially tiny defects in the workpieces, still remains a problem to be solved. In this study, we develop a hybrid framework unifying large-scale and small-scale feature extraction for single-view CT reconstruction, where two modules are designed to reconstruct the overall shape and tiny internal defects, respectively. To make the network focus on tiny defects restoration, we propose a novel FocusSSIM loss function and use an X-ray projection, and its gradient image as network input to facilitate CT reconstruction. We establish a real piston dataset containing different defects and validate that the proposed method can not only accurately reconstruct the overall shape from a single 2-D X-ray image but also restore the tiny defects precisely. This work contributes to improving the defect detection and analysis of workpieces, further promoting the exploration and application of CT in industrial automatic detection.