Research on the Construction Method of a Drainage Pipe Defect Sample Library Based on Semantic Segmentation

Research on building a sample library for drainage pipe flaws is still lacking as of right now. There is no systematic approach research in it. Supporting the intelligent interpretation model’s training and optimization is challenging. This study utilizes semantic segmentation technology as the foundation for research pertaining to the development of a sample library for drainage pipe faults. Initially, the model’s performance is meticulously evaluated in the defective image screening stage with respect to the number of defects, image resolution, format, viewing angle, and image cleansing to ascertain the most suitable screening program. Secondly, the precision of fault labeling is enhanced through the refinement of the labeling technique. This subsequently improves the model’s recognition capability. Subsequently, modulation experiments are implemented with respect to the image’s saturation, brightness, contrast, sharpening degree, and color temperature during the quality control phase of the defective samples to ascertain the most effective parameter combination scheme. Finally, a comprehensive approach based on semantic segmentation is suggested for building a sample library of drainage pipe flaws. Following the screening of image features, the model’s performance was initially enhanced, achieving maximum values of precision, mean average precision (mAP), and recall at 0.723, 0.597, and 0.674, respectively. After the optimization of the annotation technique, precision and mAP increased to 0.840 and 0.728, respectively, while recall was at 0.637. Ultimately, the baseline model attained peak performance across all metrics through meticulous calibration of image quality control parameters, with precision at 0.869, mAP at 0.751, and recall at 0.711. The findings indicate that this strategy markedly improves the drainage pipe defect detection algorithm’s training efficiency and identification accuracy. Consequently, conducting research on the development of a defect sample library for drain pipes utilizing semantic segmentation can markedly enhance the model’s detection accuracy and generalization capability. It is of great significance to the development of intelligent defect recognition technology.