Image Analysis-Based Defect Detection Model for Small-Diameter Pipes

Abstract

In Korea, houses are heated using the ondol method. This is a method of burying pipes that allow hot water to pass through the concrete of the indoor floor and send hot water through a boiler to warm the floor. The pipes used in this system are buried in the concrete that forms the floor, and once the concrete is constructed, the entire system must be reconstructed, even if there is a problem with part of it. Therefore, there are methods to inspect pipes to check for leaks before use; however, current methods have limitations. Most of these methods are indirect, which makes the detection of small construction defects challenging. To this end, a small device was developed in a previous study for inspecting pipes buried in concrete. The device identifies defects based on images captured by the device; however, there is a need to automate it because of the disadvantage of relying on manpower. Therefore, in this study, we developed a prototype model for automatically identifying defects in images captured by this device. To develop the model, an artificial intelligence method was selected to process the images and videos. The proposed model is divided into four classes: 'normal,' 'bent,' 'foreign object,' and 'punching,' and utilizes SqueezeNet, which shows similar performance with fewer parameters compared with AlexNet. The loss function of the model was CrossEntropyLoss, the optimizer was AdamW, and label_smoothing was set to 0.1 to prevent overfitting by smoothing the circle-hot-vector shape of the prediction distribution. A total of 1,873 image datasets were divided into training and validation datasets in a ratio of 8:2. The training results indicated an accuracy of 99.7%. The proposed model can automate the existing human inspection process, making it faster and easier to identify defects inside pipes.