Non-Contact Aggregate Gradation Detection Method And Application Module

Gradation refers to the distribution of aggregate in each particle size range.The volume ratio of aggregate in concrete and cemented sand gravel accounts is more than50%～70%,which is an important factor affecting many properties.Rational gradation plays a vital role in ensuring construction quality,improving building deformation resistance and impermeability in practical engineering.At present,the standard screening method is still used in the detection of aggregate gradation in engineering.This method tests the rationality of gradation through machine or manual screening and random sampling.Although the principle is simple and the technology is mature,the method’s implementation process is time-consuming and laborious,and makes the results lake of strong representation.Non-contact aggregate gradation is to obtain aggregate grading data through image recognition technology.In the past ten years,with the rapid development of computer technology,image recognition technology has made great progress.Although the research of traditional image segmentation algorithm for aggregate images has achieved some results,its universality is not strong and the traditional method is difficult to meet the accuracy requirements of today’s engineering.Deep learning method has high accuracy and wide application range,and performs well in image segmentation.However,there are too many kinds of deep learning models,and they are always with complex structure.How to apply it into aggregate image gradation test is also a problem to be solved.In this thesis,the aggregate is taken as the research object.Through on-site sampling and many grading image acquisition experiments,a large number of experimental data are obtained to construct the aggregate image dataset.This thesis analyzes the principles of several traditional image segmentation methods and then evaluates their practical application effects on aggregate images.This thesis detailed introduces the principle of instance segmentation model Mask R-CNN convolutional neural network and applies it to gradation images to quickly and accurately extract the geometric features on graph of aggregate.To sum up,the main research contents and achievements of this thesis are as follows:（1）All the aggregate images used in this thesis are obtained through the collection experiment.This thesis has designed and completed three aggregate grading image acquisition experiments,and collects more than 5000 images.Based on the above results,the aggregate grading image dataset and deep learning sample labeled dataset are established.Sufficient experimental images and measured data ensure the completion of this thesis.（2）This thesis comprehensively discusses the principles of several classical conventional methods and their advantages when applied to grading images.This thesis explores the combination of deep learning and aggregate image gradation detection.Through integrating the above methods,this thesis determines the technical process of image-based non-contact aggregate gradation detection method.Through the statistical analysis of the verification results,this method takes 80s～120s to process an image of about 20 megapixels,and the average error on geometric information of aggregate can be controlled below 5%.This result makes it feasible to detect aggregate gradation based on grading images.（3）Based on the image-based non-contact aggregate gradation detection method,this paper further integrates the corresponding gradation detection module.This research introduces its technical process and tests its performance on gradation detection.The verification results show that,under the conditions of various placement methods and particle size ranges,the correlation coefficients of the detected grading curves are more than 0.97,RMSE can be controlled below 8.0,which meets the expected accuracy.Finally,this thesis designs a concise application client for the result output part.