The Study On Generalization Of Image Feature Extraction

Nowadays,Artificial Intelligence(AI)has alread achieved a great success in the area of image processing.This is mainly attributed to a series of machine learning models,especially the deep neural network(DNN)models.Driven by the largescale training data,machine learning methods has overcome human on many image processing tasks through hierarchical feature extraction.Specifically,most machine learning methods are proposed based on the consistency between the training data and the testing data.However,in most real-life cases,the real-time data stream(test data)is commonly much more complicated than the training data,which poses a challenging problem to the generalization of the learning models.Driven by this challenge,this paper concentrates on three classical real cases that the training data and testing data is inconsistent and require generalization:(1)The scale of the training data is much less than the testing data.(2)The learning model tends to have over-fitting phenomenon in the training process.(3)There exists a distribution gap between the training data and testing data.Facing these three cases,this paper studies and improves the generalization during the image feature extraction process via a variety of techniques.The main contributions of this paper are summarized as follows:· Towards the case that the scale of the training data is less than the testing data,we build a novel few-shot face image recognition model based on partial differential equations.Specifically,we design a hierarchical partial differential equation model named “LD-PDE” through combining the Navier-stokes Equation and the rotational invariant operators.Meanwhile,to distinguish different locations in the operators,we design the location dependent mechanism based on the attention model.Experiments on few-shot face benchmarks show the generalization capability of LD-PDE when training samples are limited.· Towards the case that the deep learning models tend to have over-fitting phenomenon in the training process,we design an adaptive stochastic regularization method named “Rademacher Dropout”.Based on the assumption that each neuron in one layer owns its unique dropout rate,we derive and prove the generalization bound for DNN with dropout.Furthermore,we obtain a closeform solution for our dropout model through optimizing the generalization bound.Extensive experiments on five benchmarks show that the proposed Rademacher Dropout effectively prevents the overfitting and improves the generalization ability of deep models.· Towards the distribution gap between the training data and testing data,we design a task-specific adversarial framework for multi-source domain adaptation named TMDA.After analyzing and comparing the task-specific and domain-specific frameworks,we conclude that task-specific framework tends to refine the decision boundary more precisely.Meanwhile,the clustering of multiple source domains in the embedding space further improves the capability of TMDA.Experiments show that TMDA improves the generalization ability of image features from multi-source domains.To sum up,this paper attempts to build different models to improve the generalization of the process of image feature extraction from three different viewpoints including partial differential equations,stochastic regularization techniques and domain adaptation.Extensive testing results on multiple tasks with multiple benchmarks demonstrate that the proposed methods in this paper effectively improves the generalization and robustness of the image feature extraction process.