Neural Network Despeckling for Ultrasound Phased Diagnostics of Metallic Details

This paper describes the obtained original neural network architecture intended to process ultrasound images with fully developed speckle during phased diagnostics of metallic constructions and details. The proposed solution executes operations of convolution based on an autoencoder structure which is similar to the model U-Net. The suggested model incorporates a direct connection between the compressed and restored image, enabling dimensionality reduction while preserving fine details through skip connections. The results of modeling have shown that the proposed solution has an ability to both improve quality of the output image and detect real defects and filter artifacts masking for false defects.