Tomographic Object Detection by Unsupervised Clustering

This work introduces clustering to electrical tomography (ET) and first compares the effectiveness of different unsupervised clustering categories in tomographic object detection. First, the one-step linear back projection (LBP) algorithm is applied to reconstruct a rough image. Then, clustering is introduced to detect and refine the objects in the image, and a higher-quality image is obtained based on the clustering results. Six unsupervised clustering algorithms belonging to different categories, including K-means, Mini Batch K-means, Agglomerative, density peak clustering (DPC), statistical information grid (STING), gaussian mixture model (GMM), are compared from the aspects of clustering evaluation index, reconstruction quality and reconstruction efficiency. Simulation was carried out based on a capacitively coupled electrical resistance tomography (CCERT) system to collect the projection data under the cases of one to three objects to be detected in the sensing region. The research results show that the quality of object detection and reconstruction can be effectively improved by post-processing the image with clustering. It is found that for the single-object distribution, the grid-based clustering algorithm STING provides images with the highest quality. While for the multi-objects distributions (two or three objects), the Agglomerative algorithm belonging to the hierarchy clustering category shows advantage in achieving good images. Concerning the real-time performance, Agglomerative clustering algorithm is also preferred with less computational time than most of the other clustering algorithms. Therefore, among the combinations investigated in this work, LBP + Agglomerative clustering has the overall best performance in topographic object detection, in regarding of the reconstruction quality and efficiency.