Research On Sequence Image Classification Methods Based On Convolutional Neural Network

With the rapid development of the Internet and the popularity of video devices,video has become an important information carrier in today's society.Effectively analyzing the content in the video in different scenarios has great application value.Since there are a large number of sequence images in the video,if these images are analyzed independently,the associated information existing between the sequence images will be lost leading to a relatedly poor performance.In order to better use the correlated information between sequence images,we takes sequential images as the input data,and propose effective classification methods on 3D object classification task and action recognition task respectively.The main research work and results of this paper include:1.For 3D object classification task,this paper proposes a multi-view based 3D convolutional neural network,namely MV-C3 D.It combines different view images as a joint variable to learn spatial correlated features between them.Meanwhile,we demonstrate experimentally that MV-C3 D can achieve a good performance even only partial images are fed.Furthermore,the performance of this method on the 3D object classification benchmark exceeds most baseline methods which are mentioned in the paper.In addition,the method also performs well on the real-world image dataset MIRO,achieves 93.3% classification accuracy,which proves the practical application value of the method.2.For action recognition task,this paper proposes a multiple duration integration layer(MDIL),which consists of three parallel convolution operations.Each convolution kernel is set to a different size to extract different features.The MDIL in convolutional neural network make the method robust to many action dataset.This paper also proposes the DenseNet-3D model for motion recognition based on DenseNet,and embeds the multiple duration integration layer into DenseNet-3D to get MDI-3D.The experimental results show that the MDI-3D model has excellent action recognition performance.In addition,this paper also proposes a pre-training strategy for 3D convolutional neural networks.The pre-trained model can effectively reduce the training time on the target dataset,and the classification performance of the final model is better than that of model without this pre-training strategy.