Binary (1-bit) ultrasonic acquisition offers a low-cost path to embedded nondestructive evaluation, but discards amplitude information critical for energy-based imaging such as the Total Focusing Method (TFM). Recovering amplitude from binarized Full Matrix Capture (FMC) data is essential to unlock binary acquisition systems’ diagnostic potential. Prior work demonstrated feasibility using a U-Net autoencoder, but its ∼31 million parameters and ∼7 TFLOPs per inference make it unsuitable for real-time or embedded deployment.We propose MobileViT-V3-V1-FPN-PixelShuffle, pairing a MobileViT-V3-V1 backbone with a Feature Pyramid Network for multi-scale feature aggregation and a PixelShuffle module for amplitude regression. A key contribution is a local processing strategy subdividing FMC matrices into 8-element submatrices, exploiting transmit-receive locality while reducing memory and compute. Compared to U-Net, the architecture reduces parameter count by 7.6× and FLOPs by 13.8× for full-matrix input. The submatrix strategy further reduces FLOPs by 147× per inference, yielding an overall 9.2× reduction for complete FMC reconstruction.The model is trained on steel finite element simulations and experimental measurements, evaluated on unseen test blocks across three frequencies (2.25, 5, and 7.5 MHz), achieving mean Normalized Cross-Correlation (NCC) above 85%, Structural Similarity Index Measure (SSIM) above 87%, and 62% Mean Squared Error (MSE) reduction relative to U-Net. Generalization extends to copper and anisotropic titanium alloy (TA6V), confirming that the model learns general ultrasonic signal structure. Final validation on a binary hardware prototype, where binarization is performed by analog comparators rather than post-processing, confirms the practical viability of the method for real embedded acquisition systems.