| For buildings,cracks are the most common diseases,which accelerate the carbonization of concrete structure and steel corrosion,and bring security risks to the structure;On the one hand,it brings more psychological pressure to users,on the other hand,it seriously threatens the safety of people’s lives and property.In this paper,the wall cracks as the research object,the image processing and deep learning are combined,so that the computer has the ability to automatically identify cracks.First of all,during the sample collection and field detection,the fill light of flash often causes shadow,which is confused with the real crack.When the flashlight brightness is too low,the whole image is too dark,when the brightness is too high,the information loss of the crack pixels is caused.In this paper,a professional fill light multi-purpose self timer is designed.By the installation of LED lights in the mobile phone clamping device,fill light and take photos in the dark environment to obtain high-quality pictures;The operator can rotate the adjusting ring to make the telescopic rod reach the appropriate length to ensure safe walking in the dark tunnel.Secondly,in the process of sample training,combined preprocessing method is used to enhance the difference between cracks and background,so that the model has better compatibility in sample recognition;In this paper,according to the crack image classification,white wall black crack and concrete crack two types,according to the crack net website obtained crack photos,the average pixel size of its four small squares,choose to use two combined preprocessing methods for processing,this method can improve the speed of crack recognition performance significantly;Thirdly,a new algorithm for extracting the contour coordinates of the crack mask is designed;The mask coordinates were obtained according to RGB spatial model;After deleting the discrete mask coordinate points,the starting point of the contour is obtained;The multi matrix is used to traverse the crack mask in order,and then the mending points are carried out at the appropriate parts,or the point is restricted to participate in the contour search;According to the relationship between the mask coordinate points,namely the relationship between void ratio and co ownership,the contour is extracted in turn;Finally,it is determined by several decision conditions whether to continue searching or exiting operation when a closed profile is found.Fourthly,the training samples are screened by the sample screening algorithm;In the process of using the weight file in the previous stage to detect new samples,firstly,the original image collected by the device is clipped,and the sample set contains a large number of background images and a small number of crack images;According to the evaluation principles of whether the detection results include the recognition frame of the crack area,the confidence level given by the model for the recognition part,the number of mask coordinates,the distribution of projection interval after multiple rotations of mask coordinates,the aspect ratio of the all part and the every value of pixels,it is comprehensively judged whether some samples in the sample set need to be removed.Finally,a progressive learning method of crack image recognition model is proposed;The model is divided into three stages to train the crack samples: first,black lines are drawn on white paper to simulate the false crack,the contour of black lines is found by morphological method,and the crack labels are automatically made for training to generate the first level weight file;The first level weight file is used to identify the cracks on the white wall,and the crack coordinates are obtained after mask optimization.Then,the orderly crack contour and crack image information are extracted by using the 8 neighborhood of mask coordinate points and the common quantity relationship,and the crack label is automatically made to generate the second level weight file after training;Secondly,the second level weight file is used to detect concrete cracks in batches,optimize and extract the mask contour,integrate the sample information,continue to generate labels,cycle training,and finally generate the third level weight file.Finally,we continuously obtain the crack photos uploaded by users of crack website,train them regularly,and gradually improve the performance of weight files in the iterative training process;The trained mask RCNN model achieves 95.2%,83.3% and 79.2% for the three types of images respectively,which has a high recognition rate and can be used for the rapid recognition of building cracks.To sum up,the semi supervised learning method based on the crack image recognition model is used to carry out the research on building wall cracks by computer,which provides a new idea and method for deformation monitoring. |
