Industrial CT image reconstruction for faster scanning through U-Net++ with hybrid attention and loss function

To address this issue of streak artifacts and noise caused by faster scanning in this paper, this paper presents a prunable U-NET++ with hybrid attention and loss function (HAL-UNET++) deep learning network, which is a U-shaped framework-based network for industrial CT image reconstruction. In this reconstruction network, the prior knowledge that CT images are sparse in the gradient domain is utilized, and a hybrid loss function is designed to combine the L0 norm of gradient image and the image signal-to-noise ratio. This ensures that the network's learning process places greater emphasis on the noise content and sparsity of the image gradient domain. Subsequently, an adaptable convolutional block attention module has been incorporated into the shallow network layers to ensure the preservation of finer details in the CT images that are reconstructed. Finally, experimental objects comprising cables and rock samples with internal defects were used. The results indicate that the present reconstruction network, through pruning techniques, achieves rapid image reconstruction. Under the constraint of maintaining reconstruction accuracy, the pruned network demonstrates a reduction of approximately fourfold in network parameter volume compared to the unpruned network, resulting in a time saving of 0.24 seconds for the reconstruction of a single CT image.