Analysis And Research Of Deep Learning Based Fine-Grained Feature

Human beings mainly observe the environment through their eyes to obtain useful information,which is a component of human intelligence.As the eyes of computers,computer vision is using computer to understand and analyze image or video data,forming an important part of artificial intelligence.In the field of computer vision,fine-grained image analysis is a long-term basic problem,and widely exists in various practical applications,such as highly reliable identity authentication,highly accurate commodity recommendation and so on.The fine-grained image analysis task aims to analyze the sub category visual objects under the same category,but due to large differences within the same category and small differences among different categories,the fine-grained image analysis task has become a challenging problem.Therefore,it needs to obtain high distinctive fine-grained features to identify the small differences among different categories.At present,thanks to the rapid development of deep learning,current fine-grained image analysis tasks generally use deep neural network to extract fine-grained features from images,and distinguish different categories according to these fine-grained features.Some deep learning based algorithms have provided effective fine-grained feature analysis for specific image datasets,but the accuracy and universality of these methods need to be further improved.In addition,due to few training samples in the fine-grained dataset,the trained deep neural networks are overfitting,and they can not obtain satisfactory test performances.Therefore,in this dissertation,I propose a series of deep learning algorithms to improve the performances of different fine-grained image analysis tasks,and study the deep learning feature extraction algorithm based on global images and local images.The main research contents are as follows:1.Research on the fine-grained image classification based on deep learning.Previous methods have shown that locating the main objects in fine-grained images is conducive to improving the performance of fine-grained image classification tasks.Therefore,in this dissertation,a spatial and channel based attention mechanism is proposed to effectively locate main objects in fine-grained images and extract fine-grained features with high discrimination.In addition,after locating the position of the main object by the attention mechanism,a variety of attention data augmentation technology is employed to train the network model,which slows down the overfitting phenomenon of the model.In the test phase,it uses attention cropping and attention zooming to resample the test image,which further improves the test performance of the network model.2.Research on the fine-grained image retrieval based on deep learning.Fine-grained image retrieval tasks use fine-grained features to retrieve images of the same category.Therefore,it is important to consider how to extract distinguishing fine-grained features.In this dissertation,deep neural network is employed to extract the features of fine-grained images,and a piecewise cross entropy loss function is proposed to train the deep neural network.The piecewise cross entropy loss function introduces an appropriate amount of noise when training the deep neural network,which makes the output of the network model in a relatively stable state.Experiments show that the piecewise cross entropy loss function can effectively slow down the overfitting phenomenon of the network model in the training stage,thus improving the generalization of fine-grained features.3.Research on deep learning based local feature extraction algorithm.The quality of local features plays a decisive role in the performance of subsequent tasks(such as image matching,3D reconstruction and camera localization).For this reason,in view of the local feature extraction algorithm based on deep learning,this dissertation proposes a piecewise loss function with confusion.In the training stage,the proposed function reduces the overfitting phenomenon of the model to the training data by introducing a proper amount of confusion noise,so that the feature generalization performance of the model extraction becomes stronger.4.Research on high-resolution face recognition via deep learning based pore-scale facial feature.Some studies show that pores,like fingerprints and irises,are biometric features that can be used to identify human identity.However,due to the lack of large datasets of pores,there is still no pore feature based on deep learning.In order to promote the development of pore-scale facial feature,this dissertation proposes a 3D face constraint algorithm for pore-scale facial feature matching,which can effectively retain the correct number of facial pores feature matching,and then builds a large-scale pore-to-pore correspondence dataset.After building the large-scale pore-to-pore dataset,this dissertation proposes a novel Deeply Learned Pore-scale Facial Feature,which has good generalization performance and can be used for high-resolution face recognition task and twin face recognition task.