Can deep models benefit from standard preprocessing of pulsed thermography data?

As deep learning-based algorithms keep leading the data segmentation tasks, many efforts are being conducted for their applications in the processing of pulsed thermography sequences. Conventional data preprocessing methods, such as pulsed phase transformation and principal component analysis are still popular in the community as data enhancement methods to improve contrasts and signal-noise ratio in infrared images. In this contribution, the impact of applying two conventional preprocessing methods for pulsed thermography data on the performance and learning computational time of deep learning methods is evaluated. The presented deep learning methods are based on the Encoder-Decoder architectural form, specifically a U-Net and a SegNet. The results show that by applying the conventional transform, the training time for deep learning methods can be significantly reduced. Compared to the results on the original infrared data, the performance on the preprocessed data is kept relatively high, as demonstrated on polyvinyl chloride specimens.