A Dual Deep Network Based On The Improved YOLO For Fast Bridge Surface Defect Detection

As one of the key facilities of national transportation,bridges are closely related to social stability,and economic development.Surface defect detection on bridges is a critical step to ensuring bridge safety.Detecting the surface defects on bridges could monitor the whole condition of bridges timely and effectively.Then a repair performed in time will significantly increase the bridge’s longevity.Therefore,the research of automatic detecting technology possesses economic and practical values on bridge surface defects detection.However,there are various types of bridge surface defects,the identified common defect classes are spalling,exposed bar,hole and crack.In the practical application,different defects have a wide range of variations in appearance,besides,different defects generally overlap with each other,and a large number of images are acquired by bridge inspection robots,the existing algorithms cannot efficiently and precisely detect such defects.It is of great significance to propose an imagebased automatic algorithm applied in practice for fast defects detection.Therefore,it is an important maintenance task to study how to improve the detection accuracy of multiply defects and realize fast defects detection.To solve this problem,we improve the YOLO(You only look once)to enhance the performance of the network to detect multiple defects,YOLO-lump and YOLO-crack are proposed to form a dual deep network for fast bridge surface defect detection.On the one hand,the YOLO-lump can realize the detection of the lump defects(including spalling,exposed bar and hole)on larger sub-images,by employing a hybrid dilated pyramid module based on the hybrid dilated convolution and the spatial pyramid pooling to extract multi-scale features and to avoid losing local information caused by the dilated convolution.On the other hand,the YOLO-crack can realize the detection of the crack defects on smaller sub-images,by proposing a downsampling attention module that uses the 1×1convolution and the 3×3 group convolution to respectively map cross-channel correlation and spatial correlation of features,enhancing the foreground response of the crack in the downsampling stage and reducing the loss of spatial information.In addition,data augmentation based on generative adversarial network and focal loss are used to further improve the performance of YOLO-lumpy and YOLO-crack.In order to realize fast defect detection in a massive number of images with high resolution,we use network scaling and cross-stage partial connections to reduce network parameters and achieve a balance between detection accuracy and detection speed.We trained and tested the proposed algorithm on 169,621 high-resolution images(5120 × 5120pixels).Experimental results of the practical application show that the proposed algorithm can improve the detection accuracy of the bridge surface defects and realize real-time detection,the whole detection process takes only 0.995 seconds to handle a single high-resolution image.