BUBC-Net: Bottom-Up-Based CNN Model With Bounding Circle for Efficient Detection of Circular Sewer Pipe Joints

Regular surveying and efficient diagnosis of sewer pipe are very significant to extend the service life of sewer and ensure the sewer network functionality. The internal situation of sewer pipe is visually captured by closed-circuit television (CCTV). A dominant and characteristic feature of the internal imaging of circular sewer pipe is the strong, approximately circular reflection of the joints. Computer vision (CV) technologies have offered some alternatives for pipe joint detection to aid the sewer pipe diagnosis. Most of these methods, however, rely on traditional handcrafted feature techniques, which are not sufficiently accurate and fast, limiting their application in automation diagnosis of sewer pipe. Toward accurate and efficient sewer pipe joint detection, we created a labeled dataset named PipeCircle and proposed BUBC-Net, an efficient bottom-up-based convolutional neural network (CNN) method designed for detecting circular sewer pipe joints using a bounding circle representation. The network’s feature fusion capability is enhanced by incorporating a path aggregation network (PANet) and optimizing the structure of the detection head. In addition, the bounding circle is optimized for circle-shaped pipe joint. In extensive experiments, BUBC-Net demonstrated significant competitiveness after feature fusion, architecture tweaks, outperforming other conventional methods in terms of precision, recall, $F1$ -score, and latency of 90.1%, 78.0%, 83.6%, and 15.6 ms, respectively. BUBC-Net, moreover, provides more accurate localization of pipe joints compared to other CNN models using bounding boxes, with an increase of 1.07% in circle intersection over union (cIoU). The results indicate that the proposed BUBC-Net method can recognize sewer pipe joints well to aid the automation of sewer pipe diagnosis.