چكيده به لاتين
In this thesis, an attempt is made to replace the traditional methods of SAFT applied to the B-scan ultrasound images using the Pix2Pix neural network. Training this neural network requires a large dataset of B-scan images and their processed output via SAFT. One of the many issues one runs into is that ultrasound B-scan images are not publicly available. Due to the similarity of these data with ultrasound B-scans-the shapes involved are of arc-like patternsand due to the abundance of GPR images, these have been used to train the neural network. Consequently, in the latter stages of this work, three significant SAFT methods were utilized such as DAS, UTSR, and XTFM on three concrete data samples. The performance evaluation was made through various critical parameters related to the quality of an image, namely Standard Deviation, Entropy, Average Gradient, MSE, PSNR, and SSIM. Results were shown to improve the following SSIM scores: from 0.833 to 0.852 for the first dataset, slightly decreased from 0.352 to 0.348 for the second dataset, and improved from 0.783 to 0.836 for the third dataset. As a matter of fact, other quality metrics have also shown further improvements across these datasets. Besides, the processing time of the SAFT method with the neural network was much faster. First, the dataset UTSR was reduced from 4.22 to 1.38 seconds, the second dataset DAS from 10.3 to 1.29 seconds, and the third dataset XTFM from 12.5 to 1.36 seconds.
