A Lightweight Solution of Industrial Computed Tomography with Convolutional Neural Network

Guogang Zhu, Jian Fu

Abstract

As an advanced non-destructive testing and quality control technique, industrial computed tomography (ICT) has found many applications in smart manufacturing. The existing ICT devices are usually bulky and involve mass data processing and transmission. It results in a low efficiency and cannot keep pace with smart manufacturing. In this paper, with the support from Internet of things (IoT) and convolutional neural network (CNN), we proposed a lightweight solution of ICT devices for smart manufacturing. It consists of efforts from two aspects: distributed hardware allocation and data reduction. At the first aspect, ICT devices are separated into four functional units: data acquisition, cloud storage, computing center and control terminals. They are distributed and interconnected by IoT. Only the data acquisition unit still remains in the production lines. This distribution not only slims the ICT device, but also permits the share of the same functional units. At the second aspect, in the data acquisition unit, sparse sampling strategy is adopted to reduce the raw data and singular value decomposition (SVD) is used to compress these data. They are then transmitted to the cloud storage. At the computing center, an ICT image reconstruction algorithm and a CNN are applied to these compressed sparse sampling data to obtain high quality CT images. The experiments with practical ICT data have been executed to demonstrate the validity of the proposed solution. The results indicate that this solution can achieve a drastic data reduction, a storage space save and an efficiency improvement without significant image degradation. The presented work has been helpful to push the applications of ICT in smart manufacturing.

Introduction

Industrial Computed Tomography (ICT) is essential for non-destructive testing in modern manufacturing, yet traditional systems face a significant "weight" problem. These monolithic machines generate massive datasets—often exceeding 10GB per scan—creating bottlenecks in real-time production lines and overwhelming industrial networks. Reducing data through sparse sampling typically introduces severe streak artifacts and noise, rendering images useless for quality control. This paper proposes a lightweight solution that slims both hardware and data requirements through a distributed IoT architecture and advanced algorithmic compression.

Proposed Framework

The proposed solution decouples the traditional ICT system into four functional units interconnected via the Internet of Things (IoT). This distributed approach allows the bulky processing hardware to be moved off the production line.

The proposed distributed IoT architecture for lightweight ICT.
  1. Data Acquisition: The only unit remaining on the production line, focused on rapid X-ray projection.
  2. Cloud Storage: Acts as a centralized data hub for the system.
  3. Computing Center: A remote server cluster dedicated to heavy reconstruction tasks.
  4. Control Terminals: Mobile or desktop interfaces for monitoring and system management.

Data Reduction and Restoration

To address the digital weight of ICT, the framework employs a two-stage compression strategy. First, sparse sampling is used to capture only 1/12th of the standard X-ray projections. Second, Singular Value Decomposition (SVD) compresses the resulting matrices by retaining only the top 30 singular values, exploiting the inherent redundancy in scanning homogeneous industrial parts.

CNN-based Image Restoration

To mitigate artifacts caused by aggressive data reduction, a custom Convolutional Neural Network (CNN) was developed. The architecture integrates a U-Net structure for image-to-image mapping with DenseBlocks to enhance feature reuse while maintaining a lightweight model profile.

The hybrid U-Net and DenseNet architecture used for artifact removal.

The network utilizes residual learning, focusing on identifying and subtracting artifacts from the corrupted image rather than reconstructing the entire image from scratch. This approach ensures high-fidelity restoration even from highly compressed inputs.

Experimental Results

The system was validated using practical ICT data from aluminum automotive components. The combined strategy achieved a massive 373.32x data reduction, shrinking a 9.4 GB dataset to just 0.025 GB. Despite this compression, image quality remained high:

  • Feature Similarity Index (FSIM): Improved from 0.7241 in corrupted images to 0.9258 in restored images.
  • Information Weighted SSIM (IW-SSIM): Improved from 0.4653 to 0.9006.
  • Processing Speed: The CNN restores a single slice in approximately 0.8 seconds on a mid-range GPU.

Conclusion

By distributing hardware functions and applying aggressive algorithmic data reduction paired with CNN-based restoration, this framework provides a viable path for integrating high-resolution ICT into smart manufacturing environments. The solution significantly improves efficiency and reduces storage costs without compromising the diagnostic quality required for industrial inspection.

Summary updated

Abstract

As an advanced non-destructive testing and quality control technique, industrial computed tomography (ICT) has found many applications in smart manufacturing. The existing ICT devices are usually bulky and involve mass data processing and transmission. It results in a low efficiency and cannot keep pace with smart manufacturing. In this paper, with the support from Internet of things (IoT) and convolutional neural network (CNN), we proposed a lightweight solution of ICT devices for smart manufacturing. It consists of efforts from two aspects: distributed hardware allocation and data reduction. At the first aspect, ICT devices are separated into four functional units: data acquisition, cloud storage, computing center and control terminals. They are distributed and interconnected by IoT. Only the data acquisition unit still remains in the production lines. This distribution not only slims the ICT device, but also permits the share of the same functional units. At the second aspect, in the data acquisition unit, sparse sampling strategy is adopted to reduce the raw data and singular value decomposition (SVD) is used to compress these data. They are then transmitted to the cloud storage. At the computing center, an ICT image reconstruction algorithm and a CNN are applied to these compressed sparse sampling data to obtain high quality CT images. The experiments with practical ICT data have been executed to demonstrate the validity of the proposed solution. The results indicate that this solution can achieve a drastic data reduction, a storage space save and an efficiency improvement without significant image degradation. The presented work has been helpful to push the applications of ICT in smart manufacturing.
BibTeX
@inproceedings{zhu2020lightweight,
  author = {Guogang Zhu and Jian Fu},
  title = {{A Lightweight Solution of Industrial Computed Tomography with Convolutional Neural Network}},
  booktitle = {NDT},
  year = {2020},
  doi = {10.1016/j.ndteint.2020.102347},
}
Figure