Research On Recognition Of Aggregate Shape Features Based On Convolutional Neural Network

The volume and weight of the aggregate account for more than 90% of the asphalt mixture,which is an important part of the mixture.The shape characteristics of the aggregate affect the performance of the road by affecting the performance of all aspects of the mixture.Therefore,in-depth research should be conducted on the shape and characteristics of the aggregate.In recent years,Convolutional Neural Network(CNN)has become a new research direction in many disciplines.It also shows its superiority in the field of graphics processing and recognition,playing an increasingly important role.Convolutional neural networks are implemented to classify and identify the twodimensional shape features of aggregates.Introduced a new method to study the shape of aggregates and classify according to aggregate shape characteristics.It is important to evaluate whether the processing of coarse aggregates meets the requirements,to guide the adjustment of aggregate gradation,controllability,digitization and precision.Significance;It is also of great significance for a better study of aggregate shape characteristics and its impact on asphalt mixture road performance.The main idea of the thesis is to select the shape index as the basis for the classification of the aggregate shape characteristics based on the research on the shape characteristics of the aggregate,and to divide the aggregate into 4 categories based on the results of the shape index of the regular polygons.According to the aggregate shape classification results and aggregate information extraction requirements,a seven-layer aggregate shape feature recognition convolutional neural network including two convolutional layers,two pooling layers,two fully connected layers and a softmax layer is designed.Apply it to the two-dimensional morphological classification of aggregates.By adjusting the network parameters,the evaluation indexes are used to discuss the recognition effect of each network model on the aggregate shape characteristics.Then,the second classification of the aggregate is carried out,and the angletexture coefficient of the aggregate is selected as the new aggregate shape classification index to perform the three classification of the aggregate to determine whether the aggregate shape feature recognition model can be widely used in the aggregate shape classification.The main research content and results include the following aspects:1.According to the aggregate shape characteristics and its classification objectives,it is proposed to use a seven-layer convolutional neural network containing two convolutional layers,two pooling layers,two fully connected layers,and a softmax layer to solve the aggregate shape feature extraction And classification issues.2.On the basis of the network model A,by changing the learning rate,the number of training pictures per batch,the size and number of convolution kernels,the number of network layers,the proportion of the verification set in the training set and other parameters,use the training accuracy,verification accuracy,The five indicators of overfitting rate,verification loss,and training time evaluate each network model.It is concluded that the learning rate is 0.0001,the number of training pictures per batch is 8 or 128,the size of the convolution kernel is 3×3,the number of convolution kernels is 16 or more,and the network depth is 2(the network model has two Convolutional layer,two pooling layers),the verification set accounts for the training set 0.2(the network model is used to train more than or equal to 1920 pictures),the training accuracy and verification accuracy of the network model can reach 99.22% and 94.38%,respectively.the above.3.From the test results of 120 aggregates randomly tested on the test set,it is concluded that the probability of the aggregate classification result of the test is more than 95%,and the probability of the aggregate classification result of the test is more than 90%,73%,The classification result probability is more than 85% accounted for 83%.The recall rate R of the aggregate shape recognition model is 90.83%.4.In this paper,the shape index F of the aggregate and the texture-parameter AT of the aggregate are selected to perform four classifications,two classifications and three classifications on the same batch of aggregates,and the aggregate shape feature recognition model based on the convolution neural network is used for training.Judging from the training results and test results of the model,it can be applied to the classification of aggregate shape.Therefore,this paper introduces a new method for researching the shape of aggregates and classifying the characteristics of aggregates.Its advantage is that it can automatically read the aggregate photos and output the classification results with high accuracy without manual intervention.