Image Classification And Object Localization Based On Deep Learning

Computer vision is a hot research area nowadays.Recognition and analysis of massive image data can be widely used in security,education,medical care and transportation.In recent years,with the increase of image data and the progress of hardware computing capability,Deep Learning(DL)technology,especially Convolutional Neural Network(CNN)has been widely used in the field of computer vision,where it has achieved outstanding performance.Image classification and object localization are two basic tasks in computer vision.Currently the most widely used method is deep convolutional neural network.In the classification task,increasing network's depth can reach better performance but lead to great computational cost at the same time,and it is also an issue to control the parameter size effectively..Object localization is a further understanding of the images.Beyond image recognition,we can continue to mine the location information of a specific object,which helps to express semantic contents of the visual scenes more completely.Researchers focus on constructing effective image features for more accurate object location.This paper mainly studies image classification,object location and the applications of these two tasks.The contributions of the paper are listed as follows:1.Based on the residual network,an hourglass structure is proposed and embedded into the residual network as a branch,which can fully preserve the shallow layers' information.This structure can make the low-semantic and high-resolution features of the bottom layers control the high-semantic and low-resolution features of the top layers.Experimental results on both CIFAR-10 and CIFAR-100 datasets proves the effectiveness and robustness of the proposed method.2.Introducing contextual information into the Faster R-CNN detection model to assist in the classification and location of candidate regions is proposed.By modifying the Regions of Interest Pooling(RoI Pooling)algorithm,the contextual information of the candidate regions and the global information of the whole image are merged with the original features.Experiments on Pascal VOC datasets show it can improve the detection accuracy of larger objects;3.This paper applies the hourglass network and the localization structure with contextual information to ImageNet contests.In the classification task,the integration of the bottom layers with the top layers is simplified,where less parameters bring faster convergence and better classification performance.In addition,data pre-processing and activation function are involved.In the localization task,we firstly confirm the output number of the network,and then combine the contextual information and global information to locate the objects more precisely and achieve better results;4.This paper studies logo detection in natural scenes.First,we augment the FlickrLogo-32 datasets by manual collection and annotation.Second,several logo detection networks based on different.backbones are constructed and trained by transfer learning.Finally,K-means clustering algorithm is used to set more reasonable hyper-parameters so that the Region Proposal Network can generate more accurate candidate boxes and improve the accuracy of logo detection.