Deep Learning-Based Defect Detection From Sequences of Ultrasonic B-Scans

Ultrasonic testing (UT) is one of the commonly used non-destructive testing (NDT) techniques for material evaluation and defect detection. The acquisition of UT data is largely performed automatically by using various robotic manipulators which can ensure the consistency of the recorded data. On the other hand, complete analysis of the acquired data is still performed manually by trained personnel. This makes the reliability of defect detection highly dependent on humans' knowledge and experience. Most of the previous attempts for automated defect detection from UT data analyze individual A-scans. In such cases, valuable information present in the surrounding A-scans remains unused and limits the performance of such methods. The situation is better if a B-scan is used as an input, especially if the dataset is large enough to train a deep learning object detector. However, if each of the B-scans is analyzed individually, as it was done so far in the literature, there is still valuable information left in the surrounding B-scans that could be used to improve the precision. We showed that expanding the input layer of an existing method will not lead to an improvement and that a more complex approach is needed in order to effectively use information from neighboring B-scans. We propose two approaches based on high-dimensional feature maps merging. We showed that proposed models improve mean average precision (mAP) compared to the previous state-of-the-art model by 2% for input resolutions of $512\times 512$ pixels, and 3.4% for input resolutions of $384\times 384$ pixels.