Research On Pavement Crack Recognition Algorithm Based On Caffe Dual Model

There is a very broad realistic meaning to replace the traditional manual pavement detection which requires a log of manpower,financial resources and time with the pavement crack detection based on the technology of computer image recognition.This paper first demonstrates the research findings of related domestic and international study of pavement crack detection,then introduces the application of convolutional neural networks on image recognition and establishes a basic flow framework for recognition algorithms.Meanwhile,by choosing and labeling appropriate pavement image,the pavement image database is constructed.Then pretreatments are carried on the database in order to enhance the quality of the images by highlighting the detail characteristics of the crack and remove the interference factor such as environmental noise.The schemes I use here includes image cutting,image filtering and denoising,image gray-scale averaging,image binarization,morphological image erosion,and image normalization.Experiments and practices are carried on to prove the availability of these schemes.After that,the preprocessed output image is input into the convolutional neural network to train the recognition model of the road surface image.And through the experimental data,I am able to optimize the model structure,improve the accuracy,and prevent the model from appearing overfitting.Finally,a convolutional neural network structure with excellent recognition effect is obtained.Then,in view of the mistaken image of the pavement,the characteristics of the pavement image are observed by the output.And compare the test data of the recognition model constructed by different input images.The recognition model constructed by grayscale averaging image is found to have good anti-interference ability,better grasp the details of cracks,and identify road surface images with interference items such as road shadows and brake marks.The recognition model constructed with two valued images has a better ability to recognize the road images without interference.In order to retain the advantages of the two models at the same time,a pavement crack recognition algorithm based on Caffe dual model is designed and implemented.The experiment proves that the complementary advantages of the two models can effectively improve the recognition performance of the final model.In order to further strengthen the model constructed with gray mean image,the ability to recognize the road image with interference items is identified.Aiming at its recognition characteristics,we redesigned a modified convolution neural network structure,deepened its network level,and added multi-scale feature extraction,so that the recognition model constructed by gray mean image has stronger anti-interference ability.Finally,the experimental comparison shows that the dual model algorithm has better anti-interference ability and recognition accuracy,and the recognition model with a recognition rate of over ninety-seven percent is constructed.