Convolutional Neural Networks And Its Applications

In recent years,Convolutional Neural Network(CNN)has made great progress in the field of computer vision.But the mathematical model has not been fully ver-ified and explained.Therefore,analyzing theoretically the approximation ability and generalization performance of CNN mathematical model has important research value.This paper essentially analyzes the internal mechanism of deep CNN model over shal-low CNN model,and further investigates the approximation ability and generalization performance of CNN model.Theoretical research results show that deep CNN is con-sistent in the continuous function space C((?))defined on any tight subset (?)(?)R~d and the generalization error can be bounded by the sample size m,the number of layers J and the number of adjustable parameters w in CNN.It provides theoretical explanation for CNN model.In addition,with the rapid development of science and technology,the ability of acquiring data is becoming stronger in various disciplines,and many problems can be transformed into extracting objects of interest from images.Therefore,how to apply CNN in these problems reasonably has great significance for promoting the improve-ments.In the practical applications,we propose multiple automated workflows aiming at different tasks.(1)Aiming at extracting mitochondria from large-scale electron microscopy(EM)data,we propose a method based on 3D supervised full convolution network to recon-struct mitochondria.The 3D convolution is helpful to obtain the spatial information of mitochondria,and the deep supervision is helpful to avoid gradient vanishing in training process.They help to achieve accurate segmentation and reconstruction.The experi-mental results on two EM datasets demonstrate the effectiveness of proposed method.(2)Aiming at high-throughput Ni-based superalloys,we adopt the image process-ing methods to map the composition,microstructure,and hardness in the full EM image and establish the correspondence between material composition,microstructure,and property.In addition,a U-Net model is used for microstructure identification and a set of morphological parameter extraction methods are established.This provides a lot of material data for the research of Ni-based superalloys.(3)Aiming at mouse home-cage recording videos,we establish an automatic anal-ysis method for mouse home-cage behavior,including location and track of mouse home-cage based on U-Net,motor status analysis and Electroencephalography(EEG)classification.Automated analysis of mouse home-cage behavior shows the high cor-relation between the mouse behavior and EEG characteristics.The statistical results indicate that behavioral transitions occur not randomly but primarily between neigh-boring states.It reveals the transition law of mouse behavioral states.(4)Aiming at Microwave Radiation Imager(MWRI)onboard the Fengyun me-teorological satellite,we propose a new method to estimate the geolocation errors in MWRI data.In the coastline detection,we establish a surface fitting interpolation mod-el by involving more observations,and extract the line with the large gradient as the coastline.In the geolocation error measurement,we employ the iterative closest point algorithm to determine the correspondences between the detected coastline and the actual coastline.The experimental results demonstrate the effectiveness of proposed method.