| Macrostructures of zinc oxide nanorod array have broad application prospects in the fields of energy,environmental protection,and sensing.A large number of experimental studies have shown that the surface morphology and monomer structure of zinc oxide nanorods have a significant effect on their physical and chemical properties.However,due to their complexity,there is currently no effic ient method to measure the basic structure characteristics of zinc oxide nanorods automatically in large-scale,which greatly hinder the use of theoretical methods to deal with practical problems quantitatively.At present,the main method to characterize the surface morphology and basic structural characteristics of zinc oxide nanomaterials is the scanning electron microscope secondary electron imaging.The application of image processing technology at home and abroad in nanostructure electronic imaging is mainly aimed at relatively simple nanostructured monomers such as well dispersed nanopartic les.For the macroscopic body of zinc oxide nanorods,the field of electron microscopy usually contains a large number of complex overlapping projections of monomers.With a large number of features and complex shapes,it is difficult to effectively solve the problem of feature segmentation and batch measurement through traditional image processing techniques.To this end,this paper proposes to automatically identify the electronic image of cylindrical zinc oxide nanorods b ased on deep learning technology.Most of the shapes of zinc oxide nanorods are not highly regular.The objects identified in this paper are cylindrical zinc oxide nanorods with smooth curved sides.The main innovative research work is as follows:1.Construction of a cylindrical zinc oxide nanomaterial Electronic image Data set.The reliability of the data is mainly determined by the SEM imaging technology.Therefore,the validity problem is the main bottleneck in the construction of the zinc oxide nanorods electron microscopy data set.Specifically,it inc ludes three elements: first,the sample size of the data set must be large enough;second,the length,pitch,diameter,and orientation of the macroscopic volume of the cylindrical array are large enough;and third,the background outside the target feature is suffic iently diverse in order to deal with the actual s ituation that other types of nanostructures often appear.The validity of the data set determines the generalization performance of the model in actual deployment.To build a data set naturally requires a large number of effective samples.The traditional manual preparation method is not only ineffic ient,but also because the deposition effect is extremely sensitive to the hydrochemical deposition process parameters,and due to the complexity of the parameters and the inevitable human error,it will take more time to adjust the parameters correctly to obtain diverse effective deposition effects.Therefore,the problem of data set validity is transformed into the problem of experimental flux and the effectiveness of process parameter control.Aiming at this difficult problem,this paper prepares one-dimens ional zinc oxide nanomaterials based on zinc oxide nanostructures batch automatic deposition device(nano-structure self-evolving process machine,internal code is KQS).The KQS system contains 32 cavities,and the process parameters of each cavity are independently adjustable,and the effective experimental throughput is greatly improved.In this paper,the scanning electron microscope images of these samples are labeled,and a training set containing 177 pictures and a validation set of 45 pictures are obtained.2.Automatically and batchwise identify cylindrical one-dimensional zinc oxide nanorods and measure their length information.Aiming at the difficulty of identifying highly overlapping cylindrical zinc oxide nanorod arrays in SEM images,this paper uses a convolutional neural network model MASK R-CNN that introduces the region of intrest(ROI)mechanism to realize automatic identification of cylindrical zinc oxide nanorods and measure the length.Due to the large amount of overlapping nanostructures in the electronic images of zinc oxide nanorods,it is easy to lead to leak detections of the model.In order to reduce the leak detection rate,the model uses SOFT-NMS instead of NMS to filter out bounding boxes.In order to reduce the regression loss and prevent overfitting,the model adjusted the size of the generated bounding boxes,and added dropout regularization to the network.In order to make the model converge faster have a better performance,a transfer learning method is selected,the model is pre-trained with the COCO dataset,and the parameters are initialized with the convergent weights.The model achieved an AP of 0.831 on the task of identifying cylindrical zinc oxide nanorods,and achieved a good detection effect.The length of the zinc oxide nanorods was automatically measured according to the pixel s ize of the detection bounding boxes and the scanning electron microscope magnification. |
