Abstract: Addressing the problem that the excessive data volume of full matrix capture (FMC) leads to heavy transmission, storage, and processing burdens, resulting in low detection efficiency and poor system real-time performance, a deep learning-based total focusing method compressed sensing reconstruction imaging method was proposed. This method constructed, trained, and optimized three typical deep learning network models-1D-CNN, GoogLeNet and sparse autoencoder-for signal reconstruction. Taking the 1D-CNN model as an example, key parameters such as maximum iteration number, minimum branch, initial learning rate, and optimizer were optimized. The results showed that compared to the traditional basis pursuit (BP) reconstruction method, the optimized deep learning model reduced the percentage residual difference (P_RD) by 9%-12% on average, and the reconstruction time for one set of FMC data could be as low as 5.47 s, with reconstruction speed improved by approximately 406 times, effectively enhancing the imaging efficiency and engineering application value of the total focusing method (TFM).