Research On Image Detection Method Of Concrete Road Cracks Based On Deep Learning

The construction of highways is closely related to the development of the national economy.With vast territory,China had a total of 5.2807 million kilometers of highways by the end of 2021.Due to the long-term effects of natural climate,disasters,and load factors,the road surface gradually forms shallow cracks mainly in transverse and longitudinal directions.If not timely detected and corresponding measures taken,with the continuous impact of repeated vehicle rolling and natural climate,the cracking road surface will dynamically expand and gradually develop into wide and deep mesh cracks,directly affecting people’s driving safety and personal safety.Therefore,timely detection and repair of road cracks to eliminate safety hazards is of great significance.(1)To address the problem of inaccurate segmentation of crack images in complex backgrounds,a crack image detection and segmentation method is proposed that integrates attention mechanisms and dimension decoupling.The method is divided into two stages: the encoding stage and the decoding stage.In the channel semantic information perception module of the encoding stage,a channel attention mechanism is introduced to enhance the crack information from the channel dimension and suppress irrelevant information such as noise.Then,in the dimension decoupling module,the channel correlation and spatial correlation are decoupled to avoid interference from unnecessary channel and spatial correlation information during the decoding stage.In the decoding stage,a multi-modal feature fusion module is added to fuse with the corresponding level of the encoder,which restores the lost context information and decoupled channel and spatial correlation during upsampling.Then,in the spatial semantic information perception module,the crack edge information is obtained through global average pooling and global maximum pooling,and the spatial attention mechanism is combined to enhance the contextual information during the decoding process and improve the segmentation accuracy.Regarding the loss function,the Dice loss function,which is biased toward the difference between pixels,is used to balance the prediction deviation caused by class imbalance and locate the segmentation position more accurately,resulting in better segmentation performance.Experimental results show that the proposed method achieves more accurate crack localization and better segmentation performance,with MIOU of 83.9% and 87.2% on self-made and publicly available datasets,respectively,effectively solving the problem of crack image detection and segmentation in complex backgrounds.(2)To address the problem of weak generalization ability of existing crack image detection methods and the difficulty in detecting small cracks,a crack detection algorithm based on multi-scale feature selective fusion under a dual-branch structure is proposed.The algorithm is based on the encoder-decoder structure of the U-Net network.In the encoder,branch 1 extracts high-level abstract features of crack images by stacking 3x3 convolutional kernels and pooling operations layer by layer.Branch 2uses the Atrous Spatial Pyramid Pooling(ASPP)layer to extract crack image feature information at different scales,and the unique atrous convolution of ASPP can fully capture the edge information in crack images.The dual-branch structure is fused by the Receptive Field Selective Fusion Module(RSFM),allowing the model to learn the inter-channel correlation information between different branches and achieve the purpose of self-selective receptive fields by suppressing and enhancing different-scale features through weights.In the decoding stage,the high-level upsampling information and the low-level information corresponding to the encoder at the same scale are fused through the Multi-scale Attention Fusion Module(MAFM)under the attention mechanism to restore the lost semantic information at each level during the upsampling stage and improve the detection performance of road cracks.The experimental results show that the MIOU reached 85.8% and 87.3% respectively on the self-made dataset and the public dataset,indicating that the proposed method has higher detection accuracy and better segmentation effect for small cracks.(3)A self-made crack image detection dataset called CBC(complex background cracks)is introduced.In response to the common problem of weak model generalization ability due to simple background and single crack type in current crack image datasets,740 representative crack images were selected from a set of 20,000 crack images provided by researchers at Middle East Technical University and annotated based on criteria such as surface smoothness,multiple cracks,rough cracks,fine cracks,and blurring at the crack-background interface,resulting in the creation of the CBC dataset for training models with improved generalization ability.(4)Two algorithm models were developed to create a crack detection system.In this paper,Method 1 is suitable for segmentation detection of crack images in complex backgrounds,and Method 2 is suitable for segmentation detection of small cracks.For user convenience,the development of the concrete road crack detection system was implemented using the Tkinter toolkit.According to the needs of different crack scenes,the system allows free switching between algorithm models.The system implements functions such as previewing,detecting,and saving data of crack images,simplifying the crack detection process.