ABSTRACT: Quality monitoring of electronic chips is a key component of semiconductor manufacturing. To balance the efficiency and cost of acquiring a large number of projections in engineering applications, low-power industrial CT technology is commonly employed. The quality of the reconstructions derived from such projections with a significant amount of noise is greatly diminished, impacting subsequent detection tasks that rely on the reconstructions. However, the traditional models' excessive emphasis on spatial domain information leads to the loss of important details that are crucial for accurate interpretation and analysis in industrial applications, such as the blurring of edges caused by over-smoothing, which greatly compromises the evaluative value and reliability of the images and makes it difffcult to discern critical structural information essential for decision-making in industrial scenarios. To enhance image quality and bolster the adaptability of noise suppression in the low-power industrial CT domain, we proposed a spatial-frequency image noise suppression model (SFNS). Considering the requirements of industrial CT for the accuracy of reconstructed image structures, relying solely on spatial domain information can lead to a model that tends to ensure the compatibility of the overall texture of the image, which can easily destroy key structural details within the image. We have designed a hybrid loss that incorporates both spatial and frequency domain information. Additionally, we designed a frequency domain deviation-based weighted strategy that is incorporated into the frequency domain loss function, thereby dynamically adjusting the frequency domain loss weights. This not only reduces the negative impact of data from highly deviated areas on model optimization but also decreases the redundant use of low deviated data with less gradient information. We have demonstrated the effectiveness of the SFNS model through both qualitative and quantitative analysis.

Keywords: industrial computed tomography; low-power; noise suppression; deep learning