Research On Integrated Tunnel Image Acquisition And Typical Disease Detection

Subway is an important part of urban rail transit system,which has the characteristics of high density and high traffic volume.Under the influence of water rich soft soil layer,surrounding geology and soil structure,the lining surface of long-term running subway tunnel is prone to cracks,water leakage and other diseases.Therefore,rapid and effective detection of tunnel surface diseases is one of the important tasks of Metro intelligent operation and maintenance.With the growth of the overall construction level in China,the operating mileage of the subway is increasing.In the past,the way of inspecting the subway through manual visual inspection or carrying detection equipment is not only labor-intensive but also inefficient,which is difficult to cope with the growing inspection demand.An integrated and intelligent detection technology is urgently needed to realize automatic image acquisition and typical disease identification of tunnel lining surface,and reduce the pressure of daily operation and maintenance of Metro.Based on this,this thesis presents an integrated tunnel surface image acquisition system based on machine vision,image processing and deep learning technology and a theoretical model for intelligent identification of typical diseases.In this thesis,a highly integrated subway tunnel surface image acquisition device is designed and developed,forming an integrated acquisition module of eight cameras and light sources,which can realize the synchronous acquisition of the full view of the shield tunnel and the storage of high-definition images,and provide high-quality data support.The functional experiment and feasibility test of the system are carried out in the laboratory environment.The experiment shows that the current image acquisition device can meet the inspection speed requirements of 0-30 km/h,the image recognition accuracy is more than 0.5mm/pixel,and the overall weight is less than 30 kg,which can be handled by a single person.The high-efficiency interface design is carried out for the acquisition system,and the whole device can be flexibly installed on the hand push detection platform and the medium speed inspection robot,avoiding the repetitive design and manufacturing of the image acquisition system.Aiming at the problem of intelligent recognition of tunnel disease images with massive data,this paper proposes a hybrid model of tunnel multi-objective intelligent recognition.The algorithm can achieve high-precision disease recognition results by rough labeling,and automatically optimize the low-quality label images;At the same time,the data set quality and algorithm model are continuously improved in combination with unlabeled samples.In this study,a hybrid model is designed for complex and bad tunnel images,combining the advantages of semantic segmentation and image processing algorithms.The semantic segmentation realizes the rough extraction of tunnel cracks and the fine extraction of multi-objective.Based on the morphological characteristics of cracks,the extended threshold search algorithm and the local window extraction algorithm based on the crack skeleton are designed to reduce the missed detection of crack features,improve the detection accuracy of small cracks,and realize the automatic optimization of rough crack labels.Based on the crack image data,combined with a large number of unlabeled samples,a lightweight model is designed to realize the fine detection of tunnel crack image.The recognition algorithm model is verified by using the tunnel lining images collected.By creating a typical disease multi-scale data annotation method,a tunnel multi-objective semantic data set is established,which reduces the annotation time of a single high-resolution image from one day to more than ten minutes.The multi-target recognition experiment and generalization experiment are completed.The results show that the multi-target segmentation accuracy and intersection union ratio of the hybrid model are improved by nearly 20% compared with other semantic segmentation models and image processing algorithms,which verifies the effectiveness of the hybrid algorithm.The semi supervised experiment based on the optimized crack tags shows that the tags based on half of the original data set can achieve similar accuracy and greatly improve the detection efficiency.To meet the needs of fine detection of crack diseases,this paper designs a feature recognition and parameter calculation method,and integrates the Euclidean distance transformation model based on connected domain to achieve high-precision calculation of the width and length of crack texture.The integrated tunnel image acquisition system and the intelligent identification theory of typical diseases studied in this thesis have verified the effectiveness of the system and algorithm through the simulation tunnel test and the main line test of a metro tunnel in the south,and have engineering application value.In the future,the combination of semi supervised algorithm,deep learning model optimization and postprocessing algorithm will be further studied to reduce manual annotation as much as possible,to achieve efficient and high-quality intelligent detection of tunnel diseases.