Adaptive Semi-supervised Pixel-level Pavement Crack Detection Method Based On Multi-source Image Fusion

Driven by the rapid development of computer vision technology,pavement crack detection technology based on digital image processing and deep learning has become the focus of research in intelligent transportation.Huge data support and excellent network structure are the key to achieve high precision crack recognition at pixel level.However,in practical applications,it is difficult to obtain a large number of labeled learning samples due to factors such as the complex and variable environment for collecting pavement crack data samples,and the heavy workload of manual labeling.Moreover,the current detection algorithm based on two-dimensional or threedimensional images of pavements can hardly support efficient and accurate crack detection independently.Therefore,in view of the limitations of existing pavement crack detection technologies,this thesis takes 2D and 3D pavement data as the object,obtains high-quality data sets and high-performance segmentation models through target image preprocessing,multi-source image fusion,model optimization and selection,and combines semi supervised learning to reduce the cost of data annotation,ultimately achieving the goal of accurate and efficient identification and information extraction of pavement cracks.The main research contents are as follows:(1)Firstly,the features of 2D and 3D pavement image data are compared based on the collection of pavement data,and the common noise in pavement image data is summarized and analyzed.On this basis,different preprocessing methods are targeted for 2D and 3D images.For two-dimensional images,IGSR algorithm and gamma adaptive brightness correction algorithm are used to solve prominent shadow problems and uneven illumination problems in twodimensional images;For large scale noise and small scale texture noise in 3D data,improved radius filtering and bilateral filtering are used to preprocess 3D data.Compared with traditional preprocessing methods,the method in this thesis achieves targeted optimization of road surface images and improves image quality.(2)Aiming at the problem of poor performance in deep learning training due to the lack of detail and texture information in single source image data and unclear crack targets,this thesis adopts image fusion to achieve the fusion of multi-source image information,enriching the dimension and breadth of information in the image.Combining the characteristics of road surface images,grayscale images and depth images are selected as the source images for image fusion,and three image fusion methods,WAF,MGFF,and LP-CNN,are proposed.The research results show that the fusion method in this thesis has a more prominent ability in detail information extraction and target retention,and has a good enhancement for the defects and insufficient information problems in single source images.(3)A network model that is more suitable for pixel level recognition of pavement cracks is studied.Based on Deep Lab V3+,this thesis selects Focal Loss as a loss function for the characteristics of crack images,and uses spatial attention and channel attention mechanisms in the algorithm network to improve the network’s focus on the crack area.The comparative experiments on asphalt and cement pavement data sets show that the improved algorithm in this thesis can effectively reduce the misjudgment and missed judgment rates in crack segmentation.F1 reaches86.18% and 86.46%,and m IOU reaches 87.60% and 87.92%.The crack segmentation effect is more accurate and stable,which is better than that of the comparison model.(4)To address the problems of small training dataset of pavement distress and timeconsuming workload of pixel-level labeling,this thesis proposes a semi-supervised learning method to perform pixel-level recognition of pavement images.Using the improved Deep Lab V3+as a benchmark segmentation network,an adaptive equalization module is added to the basic Mean teacher model to balance the imbalance between the cracks and the background region in the crack image.Construct a confidence base by dynamically recording the performance of each category during training,and selectively expand and sample the data of underperforming categories based on the confidence base.The experimental results show that only 40%-50% proportional tag data is used,and the model recognition accuracy reaches the recognition accuracy of the supervised learning model trained with 100% data,which proves the progressiveness of the method selected in this thesis.