Research On Apparent Quality Detection Method Of Concrete Based On Machine Vision

The concrete frame structure is widely used in modern civil engineering projects because of its good stability and durability,and low cost of construction and maintenance.As the outermost protective material of the main structure of the building,concrete not only protects the building from external wind and snow,but also protects the steel bars inside the structure.The construction side only focuses on testing the internal performance of concrete and ignoring the inspection of the apparent forming quality of concrete in the actual construction process of the project.The traditional detection of concrete apparent forming quality is a rough judgment of concrete apparent forming quality through visual inspection.The standard of this detection method is not uniform.the randomness and subjectivity are great in the detection process,and the detection results are difficult to form a complete detection report and information,which causes great difficulties in the follow-up quality responsibility tracing and maintenance of building structures.The appearance of unqualified concrete has a great impact on the building structure,which not only reduces the strength of concrete,but also easily causes leakage,corrodes the internal steel of the structure.Sometimes it affects the durability and service life of the structure.So,it is of great practical significance to explore a more direct,comprehensive,efficient,convenient and highly automated detection method of concrete apparent forming quality.The defect images of concrete apparent quality forming in different engineering sites are collected by high pixel cell phone,which is based on image recognition technology and supported by deep learning technology.The convolution neural network is used for self-learning of the collected images to realize the detection,identification and classification of the four main types of defects of concrete,namely apparent honeycomb,pockmark,hole and joint leakage.At the same time,the detection method described is tested in the engineering field.The Main work is as follows:1.First of all,the construction and maintenance technology of concrete structure in Eastern China is described,and several defect types of concrete apparent forming in Jiangsu Province are investigated.The feasibility of convolution neural network basic algorithm is introduced.At the same time,the overall framework and process of concrete apparent quality detection method are designed.2.The images of four different types of concrete apparent quality defects are randomly collected from five engineering sites in Jiangsu Province after the design of the overall method.Then,the collected images are preprocessed to remove the redundant interference information.The image preprocessing method comprises the following steps: 1)Image brightness balancing processing based on Mask dodging to balance image brightness;2)Circular detection based on Hough transform to detect circular regions in the extracted image;3)A circular shape subtraction method based on a connected region labeling algorithm to subtract the detected circular region.The collected image of the circular tube has been removed,and the image brightness and color information get a better balance.At the same time,the number of captured images is appropriately augmented after the completion of image preprocessing.The augmentation method comprise: 1)An image data amplification method based on geometric transformation;2)Image data amplification method based on color space transformation.The number of images is expanded to facilitate the extraction and detection of image features by the subsequent convolution neural network.The images are manually labeled,and the labeled images are tested in three different convolutional neural networks after the processing of image quality and quantity.The test results show that the Faster R-CNN network structure model has the best detection results,and the accuracy is more than 60%.3.The concrete apparent forming quality is tested under the Faster R-CNN network structure model based on the above test results of different convolution neural networks.Firstly,the learning rate and the number of iterations are set,and all the processed images are regarded as a data set,which is divided into training set and testing set.The data sets are trained,recognized and classified under the given parameters.The detection and recognition rate of concrete pit defects is 28%,and the classification accuracy rate is 85% in the experimental test.The detection recognition rate of holes is 74%,and the classification accuracy rate is 94%.The detection and recognition rate of honeycomb is 60%,and the classification accuracy rate is 92%.The detection and recognition rate of joint leakage is 75%,and the classification accuracy rate is 93%.Except for the pitted surface,the detection and recognition rate of the other three defects are all over 60%,and the classification accuracy rate is over 90%.The detection results have certain reliability and practicality.4.The concrete apparent forming quality of different kinds of building main structure in a project under construction was tested in order to test the reliability and practicability of the research method in the actual project,based on the above method and combined with the actual project site.The different kinds of concrete surface quality images of the main structure of the building were collected in the same project site,and the image quality and quantity were processed according to the above method.Identification and classification are carried out on the basis of the completion of the detection.Compared with the traditional detection methods,this method is not only more convenient,simpler and faster,but also can be used to detect the apparent forming quality of the building facade and super-high concrete with the help of UAV.Although the accuracy of recognition is only about 70%,its real detection rate is far higher than the traditional manual sampling on the basis of comprehensive detection.At the same time,there are many interference factors in the images due to the complexity of the real engineering environment,which lead to more cases of misdetection classification during the detection and classification process.The detection classification rate is less than 60%.It is caused by many factors such as the characteristics of engineering environment and the shortcomings of algorithms.The direction of future improvement and exploration are put forward based on the engineering test results.