Differentiate tensor low rank soft decomposition in thermography defect detection

Composites are prone to defects in manufacture，which are to be evaluated for safety through non-destructive testing (NDT) techniques. Thermal images are acquired for NDT by using Optical pulsed thermography. Defect detection can be performed by proposing defect detection algorithms. However, due to the low resolution of defect contrast, the detection performance of the existing algorithm is still sub-optimal. In this work, a decomposition algorithm by differentiating low-rank tensors is proposed to extract weak defect information from complex thermal pattern disturbances for surface and sub-surface defect detection. The algorithm mines deep insight into the information on the differentiation of different ranks between structures from the results of Tucker decomposition to extract defect features. In particular, a probabilistic tensor model is introduced to correct potential mismatch patterns enhance defect contrast, and suppress noise and light spot interference. To verify the effectiveness and robustness of the proposed algorithm, a variety of complex composite specimens have been used for validation. The experimental results show that the proposed algorithm achieves better performance compared to the state-of-the-art algorithms especially in enhancing the defect contrast and suppressing the light spot in seven common samples. In overall, it can provide on average of approximately 15% F-score and 3 dB SNR improvement for validation.