Construction waste recycling filler is obtained from construction waste by crushing． Due to its diversified composition，It can be used for subgrade filling only after sorting． At present，manual screening method is timeconsuming． In this paper，convolution neural network was used for image analysis and the composition of regenerated packing can be obtained automatically． Firstly，a labeled dataset composed of 36000 granular images is created to train different CNN models． Among them，the user-defined resnet34 model with 40% Dropout rate performs the best，and its verification accuracy can reach 97% ． Secondly，the mass of particles was estimated based on the particle type and shape． Finally，the method proposed in this paper was compared with the manual screening method． For most recycled fillers，the quality difference is less than 2% ． This paper aims to improve the utilization of construction waste and it is of great significance to the popularization and application of construction waste backfill subgrade engineering．