Research On Road Crack Detection And Model Learning Strategy Based On Semantic Segmentatio

In the field of pavement crack detection,traditional image processing techniques suffer from low robustness and reproducibility.Although convolutional neural network(CNN)-based methods have made remarkable progress in this task,they are also insufficient in global context feature extraction.As a result,these models may be interfered by heavy noise from the pavement,generating fragmented crack curves with numerous false positives.Consequently,these two types of methods fail to meet the demand of pavement crack detection.Some methods attempt to improve the long-distance context dependency modeling ability by inserting attention mechanism-based plug-and-play modules into their baseline models.However,this improvement may put heavy pressure on memory and computing overhead,negatively impacting efficient and real-time pavement crack detection.Moreover,these increasingly sophisticated model architectures require larger scale training datasets.Besides,a systematic and comprehensive review of model training tricks is urgently needed in pavement crack detection to boost training and detection efficiency and accuracy without modifying model architecture.To address the aforementioned issues,this dissertation proposes several solutions.Firstly,the locally enhanced Transformer-based model offers a robust and reliable solution to extract pavement crack curve lines from noisy background images.Secondly,the consistency regularization-based semi-supervised model presents an effective solution to improve the performance of pavement crack detection models while reducing the reliance on annotated data.Lastly,the examination of various model training tricks provides a comprehensive understanding of their effectiveness and applicability in pavement crack detection,thus offering guidance to future research in this field.Specifically,(1)The proposed pavement crack detection model,LETNet,is a locally enhanced Transformer-based model designed to improve global context dependency and facilitate long and thin crack detection.The model features a convolution stem and a local feature enhance module to enhance the Transformer’s inductive bias and local context feature extraction.Additionally,other modules,such as a novel skip connection,an efficient up-sampling module,and a deep supervision-based defect rectification module,are proposed to boost model performance in the pavement crack detection task.Overall,the proposed model offers a comprehensive solution to address the limitations of traditional image processing techniques and CNN-based methods,and offers superior performance in detecting pavement cracks with high accuracy and efficiency.(2)Building on the LETNet model proposed above,a consistency regularization-based semi-supervised model is introduced to address the challenges of large-scale training data requirements,low training efficiency,and high computational memory costs associated with fully supervised methods and other kinds of semi-supervised techniques.Specifically,the decoder of LETNet is decoupled to introduce feature perturbation,allowing for the utilization of unlabeled training data.In addition,several other techniques are employed in this semisupervised framework,including the Cut Mix＿S data augmentation method,warming up function,and sharpening function,to further improve the performance of the model.Overall,this consistency regularization-based semi-supervised model presents a promising solution to alleviate the need for extensive training data while maintaining high detection accuracy and efficiency in pavement crack detection.(3)Deep models often incorporate various training tricks,but they are often not explicitly mentioned in manuscripts,leading to unfair comparisons with other models.Furthermore,few studies have focused on the applicability of different model training tricks in the context of pavement crack detection.To address this,the present work examines a series of deep model training tricks on pavement crack detection datasets,including data pre-processing,data augmentation,and model implementation.By examining and comparing the effectiveness of various training tricks,this work provides valuable insights into their applicability in pavement crack detection,ultimately leading to more efficient and accurate models in this field.