Rapid Detection Of Aggregate Gradation Based On Digital Image And Deep Learning Technology

In road engineering construction,aggregate gradation is the key factor that affects the strength of asphalt mixture.How to detect aggregate gradation quickly and accurately so as to achieve the purpose of improving the engineering quality is worth paying attention to.At present,the industry standard requires the use of screening method to obtain aggregate gradation.Screening method has some problems,such as complex detection process and limited detection scene.With the development of computer image processing technology,some scholars try to obtain gradation information from aggregate images by using digital image processing technology.Although digital image processing technology has mature theory,the segmentation accuracy of aggregate image segmentation can’t meet the requirements.In this thesis,digital image processing technology and deep learning technology are combined to segment the aggregate particles in the aggregate image,and the segmentation effect of different technical means is compared.The geometric features of aggregate were obtained based on the aggregate particle segmentation results.Based on the geometric characteristics of aggregate,the aggregate gradation was further obtained.The specific research contents and conclusions are as follows:Firstly,a fixed light source acquisition device was designed and implemented to reduce the impact of brightness on digital image processing technology and provide aggregate data set for deep learning algorithm.The threshold segmentation algorithm and watershed segmentation algorithm in digital image processing technology were used for particle segmentation of aggregates in different states.The results show that the threshold segmentation algorithm can not segment aggregate particles in the adhesion state.Although the watershed segmentation algorithm can segment aggregate particles in the adhesion state,the segmentation results are not satisfactory.Secondly,Deeplab V3+ and U-Net deep learning algorithms were used to train aggregate data sets respectively,in which U-Net network training MIo U(Mean Intersection over Union)improved 3.1% compared with Deeplab V3+.Compared with Deeplab V3+,U-Net network can significantly alleviate the problems of contour disconnection and independent points.Finally,this thesis determines to use the training results of U-Net deep learning network to segment the aggregate image.Then,on the basis of U-Net network prediction results,the aggregate particle geometry information of 2D images was obtained through Open CV software library.The shape of aggregate was evaluated macroscopically from three characteristic factors:angular,roundness and length-width ratio.The numerical range of characteristic factor can reflect the shape characteristics of aggregate particles.Finally,the projected area of aggregate particles was used as the conversion parameter between the image method and the screening method,and the aggregate grading data was quickly obtained by using the image method.This thesis analyzes the causes of errors in acquiring gradation by image method and designs error correction methods.Under the premise of the same source,no matter single file aggregate or mixed aggregate,the error between the modified image method grading results and the actual screening results is small,so as to achieve the purpose of obtaining the aggregate grading quickly and accurately.When different materials use image method to obtain grading data,it is necessary to correct the error ratio of each grain size.