Deep Learning Classification Of Stereo Target Based On 3D Model

Target classification,as the core of image processing,has made rapid progress in recent years.especially with the continuous progress and development of deep learning technology,The Convolutional Neural Networks has been widely used in target classification and recognition,and has achieved very significant results.But when using Convolutional Neural Networks to classify and recognize target,because of its complex network structure and large parameter,it often needs a lot of data for training and learning to get better classification and recognition effect.However,this is difficult to meet under the realistic conditions,the lack of training data will often lead to convolution of neural network cannot fully learn,which may easily lead to over-fitting.In this thesis,in order to improve the accuracy of Convolutional Neural Networks in the classification of target recognition,we proposed a kind of deep learning classification method based on data normalization,in addition,a deep learning classification method based on 3D model is also given,and the project is implemented according to the actual situation.The classification of deep learning based on data normalization the training dataset and the test dataset are normalized,which improves the accuracy of target recognition.In this method,we first normalized the training dataset and test dataset by histogram equalization algorithm based on image segmentation,then trained the convolution neural network with normalized training dataset,and classified the test dataset.Finally,we compare it with the traditional Convolutional Neural Networks and prove that it can improve the accuracy of Convolutional Neural Networks in the classification of target recognition.In order to improve the accuracy of the classification and recognition of the target by convolution neural network,we adopt the transfer learning from 3D model,through the production of training data from the simulation environment to help the Convolutional Neural Networks to achieve the image classification detection.In the experiment,we generate the simulation dataset from the single illumination and the multiple illumination simulation environment,and compare the different effects of the transfer learning from 3D model in various situations.At last,the accuracy of the classification of the stereo blocks is improved by training with a 3D model dataset and then fine-tuning with real images,and the validity of the transfer learning of 3D model is proved.Finally,the project method is implemented according to the actual situation and applied to the actual classification and identification work.