Instance-level recognition and quantification for concrete surface bughole based on deep learning

Bughole is one of the primary influences affecting the surface quality of concrete, the traditional measurement methods are carried out by manual inspection, which is considered time consuming and impractical. The previous studies used traditional CNN with bounding boxes to localize the defects, which is unable to locate defect boundaries effectively, and it is difficult to quantify defects. In order to overcome these obstacles, an instance-level recognition and quantification approach based on Mask R-CNN is proposed in this paper. A total of 428 raw images (with a resolution of 3024 x 3024 pixels) are cropped into 256 x 256 pixel images, and 2198 images containing bugholes are selected to create the datasets. Then, the architecture of the Mask R-CNN is modified, trained, validated, and tested using this datasets. Results show 90.0%, 90.8% average precision (AP) for the bounding box and mask, respectively. The accuracy of bounding box and mask are 92.2% and 92.6%, respectively. The quantification performance is evaluated by the measured value of seventy bugholes, results show that more than 68% of bugholes have an area error rate of less than 10%, more than 74% of bugholes have a maximum diameter error rate of less than 10%. The maximum error rate for area and maximum diameter are 22.58% and 17.58%, respectively, and the minimum error rates are 0.23% and 0.16%, respectively.