Predicting the acoustic influence map of the ultrasonic total focusing method using a depth-conditional diffusion model

The Acoustic Influence Map (AIM) shows the acoustic energy distribution of total focusing images at different spatial positions. To address the computational complexity and low efficiency of existing AIMs, a deep-conditional diffusion model is proposed. An improved U-Net is adopted as the backbone network, and a four-level encoder-decoder structure is constructed. The input channels are expanded to 34, and discrete depths are converted into 32-channel feature maps to match the image dimensions by a learnable embedding layer. The mask-depth embedding dual-condition input mechanism is realised by a single-channel mask image and a noise image. The generated images are evaluated using the Structural Similarity Index (SSIM) and Pearson Correlation Coefficient (CC). The range of SSIM value is from 0.86 to 0.93, indicating that the model can effectively preserve the details of the images. The range of CC value is from 0.68 to 0.86, demonstrating that the model is able to reconstruct the geometric structure of total focusing images. The normalised amplitudes predicted by the AIM are compared with the total focusing images of holes. The results show a correlation coefficient of 0.96 between the two datasets, indicating a strong positive correlation between the predicted values of the acoustic influence map and the actual imaging results.