Research On Image Classification Based On Deep Learning

Images are the visual basics for people to perceive the environment.People can get important information from the outside world according to images.So it is of great significance to make machines automatically complete the image recognition and classification.Feature extraction is the core part of image classification.It is very important to research efficient algorithm for extracting features in the field of image classification.Deep learning algorithm has made a breakthrough in the field of image classification in recent years.It establishes and simulates the hierarchical structure of human brain,processing data and extracting features from low to high level of the image,speech,text and other input data.So it can get close to object high-level semantic features.Therefore,deep learning has widely applied in the field of image classification,especially Convolution neural network which is a high recognition rate deep learning algorithm.The advantage of CNN is that it can make convolution computation with pixels to extract features.And the weights sharing and pooling layer in CNN reduce the number of network parameters largely,which simplify the network structure and improve the learning efficiency.As the current mainstream deep learning framework,Caffe has been widely used in industry and academia.In this thesis,the author uses Caffe framework to train and predict image datasets and uses Caffe's python interface to visualize the characteristics of the convolution layer and the outputs of the full connection layer,and then does some analyses and studies.Siamese network is a twin neural network structure,which has two identical neural networks that share the weight.So the input must be a pair of samples,either the positive samples that belong to same category,or negative samples that belong to different categories.It can map image information to low-dimensional feature space,so Siamese network can also be used for dimensionality.Based on its contrastive loss function,this thesis makes two improvements to the input of the sample.The first is to reorganize the input samples.The method of reorganization is that,try to find the positive samples with the furthest distance or the negative samples with smallest distance in a group of batch stochastic gradient descent(SGD),and then combine them together so that each sample pair can play its greatest role.The second is to discard the meaningless input sample pairs.After training for some time,the feature distance of some negative samples has exceeded the threshold margin,so it can be dropped to avoid making it meaninglessly into the network.In this thesis,the author uses Caffe to complete the code of the two innovative points and analyzes the experimental results.In the experiment,the author uses MNIST handwritten datasets and maps their features to the plane space using Siamese network,so we can make intuitive data distribution display.For the evaluation of the experimental results,the author uses the Test loss curve and the Accuracy curve as the evaluation method to evaluate the original method and the improved method quantitatively.Experiments prove that the Siamese network are improved through the two improvements in accuracy and speed respectively.