Optimization algorithm of manhole recognition based on YOLOv2

Abstract

Aiming at the current situation that there is a lack of methods other than manual investigation when there are problems such as settlement, damage, and missing of manholes, a custom YOLOv2 network model algorithm based on the ResNet-50 feature extraction network was proposed. The original algorithm is optimized from the aspects of detection classes, learning methods, pre-training model, anchor boxes’ estimation and parameter configuration. The pre-trained convolutional neural network ResNet-50 was uesd as the feature extraction network combined with the YOLOv2 original network to create a detection network, and the preprocessed training set data was trained to obtain target detector. By running the target detector on the input test set data, the detection of manholes is realized. Compared with the original YOLOv2 algorithm, the training time is respectively shortened by 47%, the recall rate and F1 are increased by 9 times and 5 times, and the accuracy and detection scores are respectively maintained at 98% and 50%. The improved algorithm can detect manholes efficiently and accurately in reality.