The combination of Full Matrix Capture (FMC) and the Total Focusing Method is nowadays considered the gold standard in ultrasonic imaging. However, with FMC, the size of the acquire data increases quadratically with the number of elements, leading to significant storage and processing requirements that limit its use in portable systems and complicates the adoption of 2-D matrix arrays. Our approach is to propose a simplified, easily scalable, and low-cost binary acquisition hardware architecture. This strategy enhances the detection of small diffractive targets (e. g. , high temperature hydrogen attack and macrotextured regions) but causes significant information loss in amplitude and A-scan interpretation. Recent work has shown that U-NET autoencoders can reconstruct FMC amplitudes from binarized data, but their high computational load restricts practical deployment. We propose a lightweight alternative based on the MobileViTV3V1FPN architecture, combining convolutional efficiency with the generalization capability of vision transformers. The network is trained on a hybrid dataset comprising finite element simulations and experiments from realistic defects. Results demonstrate that the proposed model accurately reconstructs FMC and TFM amplitudes with a low computational cost, paving the way for compact low-cost multichannel acquisition systems. Work supported by Evident Scientific and NSERC.
Niddam et al. (Wed,) studied this question.