A Study On Object Tracking And Image Retrieval Using Deep Convolutional Neural Networks

In recent years,visual object tracking and image retrieval have become very popular research fields due to the huge development demand for image semantic understanding,big data pro-cessing,intelligent monitoring analysis,human-computer interaction processing,intelligent transportation processing,motion analysis and so on.Among them,visual object tracking is one of the most popular and challenging hot research projects in computer vision.Its main task is to track the significant target in a given video sequence continuously,which also provides a good information basis for the following analysis of the target.Image retrieval is an extension of the widely-used text retrieval.Its main task is to directly use the image information to find similar images on the Internet or in the database,but this project still faces a lot of challenges,such as the scale changes,changes of color,viewing angular vari-ation,and so on.Fortunately,with the rapid development of computer capabilities and the establishment of a variety of large image databases,deep learning,especially convolutional neural networks(CNN)has been widely applied in computer vision field due to its powerful learning and strong feature extraction capabilities,and have achieved many excellent results such as object detection,image semantic segmentation,image object classification,and so on.Based on the research background of the predecessors,this thesis mainly focuses on visual tracking and image retrieval problems.At the same time,it also focuses on how to apply deep learning better on these two fields.The main work of this thesis is mainly in the following aspects:1.A new artificial features for dealing with occlusion problem is proposed in visual object tracking algorithm.In recent years,the visual tracking algorithm called CSK has a good accuracy and speed performance.It mainly utilizes the cyclic displacement property of Fourier transform to sample in the search domain cyclically,thereby it increase the number of samples so that the classifier is sufficiently trained.However,CSK directly uses the original image as the input,so there is still a lot of space for improvement in accuracy.The concept of complex form is introduced in the new artificial features extraction of the proposed.It takes into account the spatial relationship between the target and the background and the temporal relationship between adjacent sequence frames,which makes the tracking algorithm more robust for tracking the target with occlusion.At the same time,we use the entropy ratio between the target and the search domain to find an appropriate search domain size for different target sizes in different videos.Finally,since the size of the target in the video sequence may change,we use the property of the gaussian function to estimate the target size.2.Further introduce deep learning into visual object tracking problems.On the basis of VGG,using features of different layers of deep convolutional neural network with different semantic information to make a features fusion.At the same time,using video instead of images as a training set.Through the network,the original target feature is used as a template to perform normalized cross-correlation match on features of the search domain.Finally,the location with the largest response value is the new location of the target.3.A feature fusion algorithm for deep hashing image retrieval is proposed.It takes the property that using different scales of convolutional kernels can sample different semantic information.The network is divided into three branches in the middle to make features extraction and fusion,and an end-to-end relationship of image features and hash codes is established.At the same time,for the gradient disappearance problem of the sigmoid func-tion in the network training process,the sigmoid function is revised and the penalty function is added to the loss function so that the network can be better trained.Experimental result-s show that this algorithm still has good performance compared with the current excellent image retrieval algorithm.