Design And Implementation Of A Defect Detection Method For Bridge Surface Based On Semantic Segmentation

Detection of surface defects on highway bridges is an indispensable and important task in transportation infrastructure maintenance,and it is also one of the most concerned and sensitive issues for construction units at all levels and the entire society.Semantic segmentation technology can locate and segment defects on the surface of highway bridges,assist relevant departments in timely detecting safety hazards such as deformation and damage of highway bridges,and take remedial measures to effectively prevent bridge safety accidents.However,most existing methods for detecting surface defects on bridges are aimed at single defect-class images,and the detection performance is poor when the proportion of foreground objects in the image is small.To address the above issues,this paper proposes a multi-stage surface defect detection method for bridge structures based on block-based semantic segmentation using deep learning.The method is designed according to the engineering detection process for surface defects on bridges,and combines image semantic segmentation technology.Firstly,the original image is uniformly partitioned into fixed-sized small blocks.Secondly,a binary semantic segmentation network is used to filter out all background small blocks that do not contain defect areas,resulting in foreground small blocks that contain defect areas.These foreground blocks are used to construct a multi-class semantic segmentation candidate set.Finally,the candidate set is subjected to multi-class semantic segmentation to detect five common surface defects on bridges,including honeycombs,cracks,and seepage,at the pixel level.This method not only solves the problem of foreground-background class imbalance,but also improves the current situation where existing models have poor precision in locating and segmenting surface defects on bridges.The specific work contents and innovative points are as follows:(1)To address the problem of insufficient defect types in current open datasets for bridge surface detection,a custom image dataset for semantic segmentation of bridge surface defects was created.This self-made dataset contains five categories of defect data,including honeycombs,cracks,seepage,repaired and spalling,which are relevant to practical engineering scenarios and semantic segmentation requirements.Additionally,the original images were uniformly partitioned into fixed-sized small blocks to provide data support for the training of the multi-stage bridge surface defect semantic segmentation network,which helps to alleviate the problem of foreground-background class imbalance.(2)For the binary semantic segmentation stage,a method based on the dilated spatial pyramid pooling module is proposed to address the problem of foreground-background class imbalance and the single scale of convolutional receptive fields.By performing category detection and semantic segmentation on foreground defect class images and background images without defects,a multi-class semantic segmentation candidate set is constructed.Firstly,this method uses an encoder composed of multiple convolutional and pooling layers to extract and compress features from small block images.Secondly,through the dilated spatial pyramid pooling module,the encoded small block images are mapped to a multi-level feature space to construct multi-scale features and enhance the reliability of feature representation.Finally,based on the pooling index,the relevant data is placed back to the corresponding position,enabling the up-sampling of small block images to the original size.Meanwhile,the Focal Loss function is introduced to adjust the penalty weights of different categories and alleviate the negative effects of class imbalance.(3)Based on the construction of a multi-class semantic segmentation candidate set using foreground defect images,a multi-class bridge surface defect semantic segmentation method based on attention mechanism is proposed,which performs category detection and semantic segmentation on foreground defect images.Firstly,the method constructs a semantic segmentation network with an encoder-decoder structure based on skip connections.Secondly,an attention mechanism reinforcement learning module is designed and embedded into the semantic segmentation network to enhance the model’s capture ability of defect areas in small image blocks.Finally,the Lovász-Softmax Loss function is introduced to optimize the semantic segmentation model and enhance the detection ability of bridge surface defects.Based on this,we compared and analyzed our method with current mainstream models using the self-made dataset.We also conducted module ablation and data generalization experiments to further validate the effectiveness of our method.Finally,according to the needs of practical engineering scenarios,we developed a bridge surface defect detection system.