Research On Crack Recognition Model Based On Deep Learning

Crack recognition based on deep learning is an important task in computer vision,which is widely used in industry,such as crack recognition in bridges,buildings,large vehicles,and other environments.Cracks generally exist in most kinds of environments,but the number of cracks in the same environment is generally small,and the forms of cracks are diverse.In addition,crack recognition is mainly applied to robots and other similar products.This paper mainly solves the problem of insufficient computing power of edge equipment crack identification task,constructs FRCRU algorithm to solve the problem of data imbalance,and designs a general large model as technical reserve.In this paper,three challenges of crack recoginition are studied:(1)In the processing of crack data,the distribution of crack data is not uniform and there is a long-tail phenomenon?(2)In the model training of crack recognition,crack samples are complex,and general recognition is difficult?(3)In terms of model deployment,it is difficult for recognition models to be deployed in such edge devices as robots and unmanned aerial vehicles.Aiming at the problem of uneven crack data distribution and long tail phenomenon,a data enhancement algorithm FRCRU is proposed to deal with uneven crack data.This paper solved the problem of invalidation of evaluation indexes by means of data set reconstruction and introduction of new evaluation indexes.According to the characteristics of data set imbalance and crack images,the corresponding data enhancement methods are proposed in this paper,and the effective part of these methods is verified by experiments.After the effective methods are combined to form a data enhancement algorithm,the algorithm is named FRCRU and the effectiveness of each data enhancement method in FRCRU algorithm is proved by ablation experiment.Aiming at the difficulty of crack recognition,this paper proposes an integration algorithm AWS based on grid search.The model of crack recognition is studied in this paper.Firstly,Transformer models are introduced into the crack recognition task and Swin Transformer is proved to be the best model through experiments.However,a single model cannot completely solve the recognition problem.Then,this paper proposes an integration algorithm Auto Weighted Sum(AWS)based on grid search.AWS solves the problem that adjustment weights in traditional model integration cannot achieve the best effect and need manual adjustment many times by using a random grid search to verify the weights of each model on the set and adding them together.Then this paper proves the validity of the AWS algorithm through experiments.In order to solve the problem that the recognition model is difficult to deploy on edge devices,lightweight models based on Point Wise convolution are proposed in this paper.This paper studies the compression of the model.Firstly,this paper introduces an integrated model distillation method to distill the lightweight models,which indirectly completes the compression of the model.To solve the problem of the poor distillation effect of some lightweight models,this paper proposed the lightweight models PWShuffle Net V2 and PWGhost Net based on Point Wise convolution,which replaced the linear calculation module in the model by using nonlinear Point Wise convolution.Experiments show that the distillation method and the lightweight model are effective.Compared with the original Res Net18 model,the accuracy and F1 score of the compressed lightweight model improved by 6.63% and 8.65%,the reasoning speed increased by3.15 times,and the number of parameters decreased by 32.59 times.The research work in this paper has also been applied to the column checking robot of the CSG Robot Technology Company.Compared with the original model,the accuracy of about 4% and the calculation and reasoning speed of about 3 times have been improved in the enterprise data set.