Research On Particle Size Detection Algorithm Of Cohesive Aggregate Image Based On Machine Learning

Aggregate plays the role of skeleton filling in asphalt concrete.The gradation of aggregate directly affects the cohesive force and strength of asphalt concrete.The particle size of aggregate is the basis for its gradation.The traditional aggregate particle size detection mainly relies on manual caliper measurement or vibration separation,and the automated adhesive aggregate particle size detection has become an urgent need to improve construction efficiency.Therefore,based on machine learning,the particle size detection problem in aggregate cohesion scenarios is studied in this paper.First of all,through the research and analysis of the particle size detection methods of cohesive aggregates at home and abroad in recent years,because of the various interferences and challenges of aggregate shapes,sizes,and mutual occlusion in the images of cohesive aggregates,Multi-Res Unet,a cohesive aggregate segmentation model based on Inception network and residual connection optimization is used in this paper.And through comparative experiments,it is concluded that compared with Unet model and morphological watershed algorithm,the accuracy of this model is increased by 2.47% and 25.95%,respectively.It can segment the edges of cohesive aggregates more accurately,and can effectively suppress the problem of discontinuity and over-segmentation of aggregate boundaries.Secondly,for the labeled complete aggregate particles and occluded aggregate particles extracted after segmentation,data mining method is used in this paper to construct 48 shape feature factor pairs from four aspects: basic shape,overall shape,invariant moment and relative shape.The shape of the particles is characterized by these data,normalization and feature selection are input into the integrated classifier XGBoost to classify the aggregate particles,and the hyperparameter combination of the classifier is optimized through a combination of grid search and cross-validation.After optimization,the superiority of its performance is proved by comparison with various types of classifiers.Finally,for the complete aggregate particles obtained by classification,23 particle size feature factors are extracted in this paper from both the basic particle size and the overall particle size to characterize their particle size characteristics.These characteristics are selected based on the correlation coefficient and inputted into the GP-Light GBM model which is optimized based on the Gaussian process.After training,it is compared and analyzed with the traditional single geometric particle size calculation model and other Boosting models.The analysis results show that the model in this paper not only has a higher degree of linear correlation,but also a lower error in the calculation of aggregate particle size.At the same time,it has higher efficiency.To sum up,by studying the particle size detection of the cohesive aggregate,the accurate segmentation and classification of the image of the cohesive aggregate and the accurate detection of the particle size of the complete aggregate is realized in this paper.Which provides a certain theoretical support for the automatic detection of the cohesive aggregate particle size on conveyor belt,and has important practical significance for the advancement of the automation and intelligence process in the field of aggregate production and road pavement construction.