Research On Tunnel Disease Detection Technology Based On Vision Of Line-Scan Camera

China’s highway tunnel construction quantity and operating mileage have now become the world’s largest,and their healthy operation is crucial to ensure the safety of people’s lives and property.Therefore,it is particularly important to monitor their health.After years of operation,many tunnel structures have developed various diseases such as cracks,leaks,and falling-off,which greatly reduce their durability,bearing capacity,and service life,and cause various safety accidents.However,traditional manual inspection methods often require the tunnel to be closed for operation,which is time-consuming,labor-intensive,and has a low degree of intelligence,resulting in high inspection costs and shutdown losses.Existing tunnel automatic detection technologies have the drawbacks of low adaptation speed,single type recognition,low detection efficiency,and poor robustness in practical applications.Therefore,this article conducts highway tunnel data acquisition based on line array vision and researches tunnel surface disease detection technology based on semantic segmentation models.The specific work includes the following aspects.Firstly,considering the advantages of line array vision in high-speed dynamic data acquisition,this article uses a line array camera as the image acquisition device for highway tunnel detection.At the same time,a detailed analysis was carried out on the problem of image distortion caused by the influence of vehicle speed,posture,and motion trajectory in the high-speed vehicle-mounted environment.Based on the principle of distortion formation,it was summarized into two categories: longitudinal distortion and motion distortion,and they were processed separately.For longitudinal distortion,a distortion correction method based on motion velocity was proposed,which eliminates the longitudinal distortion phenomenon of line array images by matching the camera capture frequency with the speed.In addition,a distortion correction algorithm based on motion trajectory was proposed for motion distortion,which restores the distorted image to the real image coordinate system frame by frame through affine transformation of line array images.Finally,an image distortion evaluation index based on imaging targets was proposed,and a distortion correction comparative experiment was designed based on this index.The experimental results show that the longitudinal distortion correction method and motion distortion correction algorithm proposed in this article can effectively solve the imaging distortion problem of line array cameras.Secondly,with regard to the problem of low efficiency and poor robustness caused by environmental interference,multiple types of diseases,and noise in the detection of highway tunnel disease images,this paper optimizes the accuracy and real-time performance based on the Deep Lab V3+ semantic segmentation model and proposes the G-Deep Lab V3+ model.The model introduces the lightweight network of Ghost Net V2 to improve the feature extraction backbone network of Deep Lab V3+,ensuring real-time performance while maintaining recognition accuracy.In addition,based on the DFC decoupling fully connected attention mechanism,the Decoder module in the network is optimized,and the multi-scale feature map fusion method is improved to improve the global feature perception and representation ability of the network,thereby improving the model recognition accuracy.To verify the effectiveness of the model,this paper collected16 highway tunnel disease images in Chengdu City,conducted statistical analysis on the types and features of tunnel diseases,and established a highway tunnel disease semantic segmentation dataset.Finally,the improved network was trained,tested and performance analyzed based on this dataset.The experimental results show that the improved GDeep Lab V3+ model has a substantial improvement in accuracy and real-time performance.