Deep Learning Classification Network Research And Its Application In Computer Vision

Just as the importance of vision to humans,computer vision is also an essential part of general artificial intelligence.Computer vision technology,such as image classification and target detection,plays a more and more extensive role in real life,such as intelligent transportation,intelligent medical care,autonomous driving,etc.Compared with the application effect of traditional machine learning algorithms in computer vision,in recent years,the vigorous development of deep learning methods has made breakthroughs in some research directions of computer vision.In the background,the deep learning classification network based on convolutional neural network has played an important role.This thesis studies the classification network and the main work are as follows.First of all,it briefly describes the history of development from shallow neural networks to deep learning and focuses on the problems that still exist in the history of neural networks.The principles of neural networks and convolutional neural networks are introduced.With the aid of a deep learning framework,the classical classification networks are reproduced,the design ideas and structural characteristics of each network are summarized.Then described in detail the classification of different network was applied to image classification and target detection task of experimental process and results,and use the latest classification network for target detection framework to replace the main network,improve the target detection model,improve detection accuracy.Combining the fusion structure of GoogLeNet network with the cross-layer connection of ResNet,a classification network based on channel fusion is proposed and compared with other networks in practical tasks of handwritten Chinese character recognition to verify the effectiveness of the channel fusion structures.When the deep classification network cannot be separately trained due to hardware resource problems,it is proposed to improve the classification accuracy by merging the classification results of a plurality of trained classification networks onthe data set.The multi-network fusion method includes multi-network fusion based on output weighting and multi-network fusion method based on network splicing.The output weight-based fusion method can only be used for image classification tasks,and the network splicing method can also be used for other computer visions tasks.