Monocular visual semantic understanding system for real-time internal damage detection

Abstract

To quickly and non-destructively obtain internal damage data of components, endoscope technology based on machine vision has been widely utilized. However, it is still challenging to measure the internal damage of equipment intelligently and accurately in real time, especially for the internal damage of precision equipment. Therefore, an intelligent endoscope detection system based on monocular visual semantic understanding is proposed in this paper. In this system, a self-developed microprobe structure is used to construct a multi-frame full convolutional network model through multi-scale feature coupling mechanism, which effectively overcomes the feature degradation caused by low contrast imaging and uneven illumination during internal detection. As a result, it enables the automatic identification of the target region and the high − precision measurement of regional geometric parameters. Experimental results demonstrate that the average absolute error of damage size measurement is 0.029 mm, with a standard deviation of 0.0236. The average mIoU is at least 3.3 % higher than other detection methods covered in this article, and the accuracy of damage measurement is improved by about 10 %. It can realize automatic and intelligent defect identification and measurement, and meet the requirements of real-time measurement on site.