Image Classification Based On Feature Optimization Of Deep Convolutional Network

With the rapid development of Internet technology,the amount of data generated or viewed on the Internet shows a geometric growth,which generates a lot of spam,or information that does not have usage value.Subdivided into the image domain,how to accurately recommend the image data to the user,which needs to improve the classification algorithm of the image.The image data is intuitive and can contain a lot of information,which makes the need for image data classification growing increasingly.Research in the field of image classification has become a very important branch of computer vision.In recent years,deep learning makes the classification of the image very simple,however,People have higher requirements on the accuracy and speed of classification.This paper proposes some optimization methods based on the deep network and improves the model structure for image classification problems.(1)An image classification method based on the deep network and sparse Fisher vector is proposed.This method firstly extracts image features through pre-trained convolutional neural networks,and then uses sparse Fisher vector to code the features.The encoded features are fed into the support vector machine classifier for training.This method reduces the complexity of the encoding,improves the characterizing ability of the feature,and enhances the accuracy of classification.(2)An integrated and optimized classification method for deep network features is proposed.This method firstly extracts the feature of each pooling layer in the convolutional neural network.The training parameters are the weights of a set of filters instead of the weights of the network,which can select features through the pooling layer.Pooled features in the convolutional network are all trained by such a set of filters.Finally,we uses the ensemble learning strategy to obtain the required classification tags.This method does not require retraining the network.We can greatly increase the speed and improve the accuracy by the pre-trained network parameters.(3)An image classification method based on metric learning and multi-model fusion is proposed.This method firstly uses SSD(Single Shot MultiBox Detector)for each sample data to detect target,and then crops the part where the target is located.Finally,this method uses an improved Siamese structure and multi-model fusion to process the cropped image.This method can reduce the influence of the background and overcome the limitations of the single model.and then improve the classification accuracy.