| Accurate measurement of particle field is a problem in many fields.In the study of material micro spray,material dynamic response and aerodynamics under impact loading,the particle size may reach submicron or even nanometer level,and the particle velocity can reach km/s,which requires high spatial and time resolution.Because of the advantages of fast,non-contact,high resolution and three-dimensional reconstruction,holography is widely used in particle field detection.Compared with digital holography,traditional optical holography is more suitable for particle field with large field of view,high resolution and high speed.The imaging scene of these high energy shock experiment is complex,which leads to low image quality and particles contrast,uneven distribution of particle size and density,and interference of noise and artifacts.It is a difficult problem to identify particles accurately and quickly from the particle field,determine their position and calculate the area.In this essay,we focus on the interpretation of high-speed transient 3D particle field,study to achieve accurate and efficient spatial positioning and size measurement of multi-scale particles.The large span on the spatial scale is one of the characteristics and difficulties of particle field measurement.To solve this problem,we propose a Primary-Auxiliary dual frame Network(PANet),which is composed of PrimaryNet(PNet)and AuxiliaryNet(ANet).PNet is used to complete the preliminary measurement of particle field,detect the medium and large-scale particles.ANet is applied as a supplement to detect the small-scale particles which are failed to be identified by PNet.PNet and ANet cooperate to realize the measurement of multi-scale particles.In order to reduce the parameters and complexity of the network,ANet shares feature maps of PNet.PANet uses multiscale convolution module to efficiently extract features and reduce the amount of parameters.The network introduces feature pyramid module to obtain multi-scale features and enhance the detection ability of multi-scale particles.According to the transient and randomness of high-speed particle field,a designed loss function and semi-supervised learning is used for network training.The network performance is evaluated by experiments of high-energy laser shock aluminum target and droplet breakup in high Mach shock save.The particle field data consists of continuous multi-layer images and each image corresponds to a slice of the observed particle field along the axis.The experimental results show that the proposed network can recognize,locate and segment multi-scale particles accurately and quickly.The size range of particle recognition is 2 × 2 to 30 × 30 pixels,the recognition rate is over 90%and the error of focus layer is less than 2 layers.The processing time of each 2048×2048 image is about 0.6s.Entrusted by the Institute of fluid physics,Chinese Academy of Engineering Physics,we develop a particle field interpretation software with Python to realize the interpretation,browsing and modification of particles in particle field. |
