LSIDA-YOLOV7: An Optimized YOLOv7 Based on Local Sensitive Information Data Augmentation for Sewer Pipeline Defect Detection

Abstract

Defect detection in urban sewer pipelines has become a popular task in recent years. To achieve this goal, this paper utilizes images containing 16 types of pipeline defects collected by CCTV pipeline robots under different pipe conditions, along with datasets gathered by some companies, to construct an urban sewer pipeline defect database. Data augmentation methods such as mosaic are employed to expand the dataset and prevent the defect detection algorithm from overfitting. By combining data augmentation methods for small and micro-sample defects, an efficient DA-YOLOv7 model is constructed. The optimized model achieves a mAP of 95.93%, F1 score of 94.99%, accuracy of 95.01% and the recall is 95.43%, and an average detection time of 0.025 seconds, demonstrating effective application in detecting urban sewer pipeline defects under complex conditions.