Lightweight Crack Learning Model Using Morphology for Infrared Sensor Images

Since various cracks can lead to significant loss of life and property, early detection and repair of cracks in structures is a very important procedure to prevent social losses. However, since it is difficult to detect cracks under paint or wallpaper using conventional techniques, it is necessary to use infrared-based thermal devices. In order to improve the performance of crack detection using infrared thermal imaging, we analyzed various image preprocessing methods and deep learning models. We developed a lightweight deep learning model using morphology preprocessing and found that morphology can be used to effectively highlight the boundaries and shape of the cracks. In particular, the lightweight Yolo model used with morphology had a faster processing speed and a similar level of accuracy with a smaller number of parameters. We demonstrate that lightweight deep learning can perform similarly to heavier models when appropriate image preprocessing is applied, which is useful in low-end portable devices.