APPLICABILITY OF MACHINE LEARNING FOR IMAGE RECOGNITION AND OBJECT DETECTION IN CCTV SURVEY IMAGES OF SEWER PIPES

Abstract

 In order to improve the efficiency of CCTV surveys of sewer pipe, the applicability of machine learning for image recognition and object detection was investigated using actual survey videos. For image recognition, the effect of transfer learning and generalization performance of four models, MobileNetV3-Large, ResNet-50, EfficientNet-B4, and EfficientNet-B4 Noisy Student, were compared in a six-class classification using F-means. ResNet-50 was better for the training data set without transfer learning, EfficientNet-B4 Noisy Student was better with transfer learning, and EfficientNet-B4 was better in generalization performance for both without and with transfer learning. For object detection, the mAP was 90.76% for 3-class classification, indicating its applicability.