| Image retrieval is a topic of great value in the field of computer vision,automatic image annotation is a key step in image retrieval and image understanding,and person re-identification is the specific application of image retrieval in the real pedestrian scene.In this paper,automatic image annotation and person re-identification were studied separately.This paper focus on the following issues:Training deep network prone to over-fitting,structures of traditional annotation models are cumbersome,there are less research in person re-identification based on real scene.Then a number of solutions are proposed in the aspects of data augmentation,annotation framework,practical application and so on.The main works of this paper include:(1)For the problems that deep neural network is difficult to train small-scale datasets and it is difficult for traditional data augmentation methods to extensively expand multi-label datasets.Combined with some traditional methods of data augmentation,a data augmentation method for multi-label based on Wasserstein generative adversarial networks(ML-WGAN)is proposed.The method trains the generator of WGAN to gradually approaching the data distribution of a single multi-label image,the images generated during the iteration are regarded as the supplement to the original image data.The proposed method can greatly expand the multi-label datasets,reduce the over-fitting problem in the process of training deep neural networks,and enhance the generalization performance of image recognition and annotation models from the data itself.(2)Traditional shallow machine learning algorithms lack generalization performance when dealing with complex classification problems.The traditional annotation model considers the feature extraction and classification and annotation as two independent tasks to study,resulting in the problem of complicated structure.This paper presents an automatic image annotation model based on end-to-end deep convolutional neural network(E2E-DCNN).The model firstly converts image annotation into the multi-label classification problem,and carry out self-adaptive feature learning by using deep CNN structure such as ResNet.Then the model use the multiple cross-entropy loss function to establish the end-to-end annotation structure for training.Finally,the model enhance the annotation effect through the structural optimization of deep CNN and the contribution of deep learning data augmentation methods.The model uses only a single deep CNN structure to complete the automatic multi-label image labeling task,while effectively improving the annotation effect.(3)In view of the fact that there are few researches on pedestrian image retrieval in real scene.This paper proposes a Faster-RCNN person re-identification method based on data enhancement,this approach combines deep learning data enhancement with Faster-RCNN.Firstly,the pedestrian position in the real scene is extracted through a deep CNN pre-training model,and then the distance measure of the target pedestrian is calculated using a non-parametric loss function.The method integrates pedestrian detection and pedestrian recognition into a single end-to-end framework,and effectively improve the effect of person re-identification.Experimental results on common datasets for image annotation,common datasets and real scene datasets for person re-identification show that:The proposed method effectively reduces over-fitting problems in model training,enhances the labeling effect of middle and low-frequency tags and real-time person re-identification.In summary,the method proposed in this paper can improve the recall rate of the entire model in the aspect of automatic image annotation,and they also can enhance the effect of person re-identification in the real scene in the aspect of person re-identification. |
PDF Full Text Request