Detection And Classification Of Concrete Bridge Cracks Based On Deep Learning

China is a big country with bridges and a bridge extending in all directions brings convenience to the people's work and life.It also conceals many invisible dangers.Once these hidden dangers erupt,they will inflict immeasurable damage on people's lives and property.First of all,the technical and management levels of various construction teams are uneven,and the quality of construction materials and construction technology cannot be guaranteed,making some bridge projects more or less have some diseases.Second,some irrational use of bridges,such as overweight,overload,and speeding,has accelerated the production of bridges or increased the severity of bridges.In the end,some of the bridges built in the early days had lower load standards and face severe challenges under today's traffic conditions.Therefore,the detection,prediction and evaluation of existing bridges are important challenges for the safe operation and management of bridges.To achieve scientific and effective maintenance of bridges,we must first select scientific and reasonable detection methods for the inspection and evaluation of defects in bridge structures.At present,the methods of bridge detection mainly include manual observation methods and non-destructive testing methods.The manual detection method usually requires the maintenance personnel to use a special tool to visually inspect the bottom of the bridge for damage caused by manual inspection.The detection accuracy is low,the efficiency is low,the labor intensity is high,and subjectiveness is strong.Therefore,we need a fully automatic adaptive bridge crack detection method.In view of the above situation,this article started work from the following two aspects.(1).To solve this problem,the characteristics of bridge cracks are firstly analyzed.There are various types of cracks in the bridge,noise is everywhere,the illumination is not uniform,and the proportion of pixels in the cracks is small.Then,according to the diversity of fracture morphology,a variety of features of different scales and directions are extracted.These features are relatively certainly more,and significant feature extraction is required.The image is transformed to frequency for problems such as noise and uneven illumination.In the domain,some noises are filtered out.For the problem that the percentage of pixels in the cracks is small,the region of interest is extracted from the image.Finally,a simplified pulse-coupled neural network is used for crack detection and verified by experiments.After verification,the proposed algorithm is based on The bridge crack detection method in the area of interest has a high accuracy(91.97%)and high efficiency(processing time 5.82s per image),which has strong practical significance.(2).Based on the conjecture that different bridge diseases lead to different bridge cracks,this paper proposes a rapid self-learning-based bridge crack image classification method TSLBCC(Transfer Self-Learning Bridge Crack Classification)based on self-learning and sparse coding.Complexity and training complexity are constants.Firstly,the improved main element analysis and whitening methods are used to reduce the dimension of the high-resolution and high-correlation bridge crack images to achieve accelerated convergence.Secondly,the scale-invariant features of the training set are extracted and then expressed using sparse coding.Feature dictionary;Finally,the test data set is trained to obtain the features represented by the feature dictionary,then the maximum spatial pyramid is pooled,and the pooled data is input to the linear multi-classification support vector machine for classification.It can not only detect whether there are cracks in the bridge image,but also can determine the type of cracks,evaluate the structural characteristics of the bridge cracks and the bearing capacity of the bridge.Experimental results show that compared with the traditional methods,the algorithm has higher classification accuracy and lower computational complexity.