Research On Image Recognition Methods Under Deficient Or Noisy Training Conditions

Recently,owing to deep learning,image recognition technologies have achieved many dramatic breakthroughs,however,the performance of existing algorithms and models will be influenced when they are confronted with imper-fect training sets.This thesis mainly focuses on three non-ideal training con-ditions including deficient sample labels,deficient training classes and noisy sample labels,and researches image recognition methods in aspects of multi-label learning,zero-shot learning and noise-robust learning,aiming to improve the generalization performance of classification models.The main research contents in this dissertation are as follows:For the problem of deficient sample labels,the existing classification-based methods are likely to ignore the relationships between tag semantics and image features,and the other probability-based methods lack of effective utilization of prior annotation information.Thus,this dissertation proposes a multi-label learning method based on precisely constrained matrix completion,which formulates the task of multi-label learning as the optimization of the image-tag matrix and introduces constraint to preserve initial correct annota-tion as precisely as possible.The optimization objective is to minimize the discrepancy of relationships between image visual features and label semantic information,which is solved through the linear alternating direction method.The effectiveness of the proposed method has been demonstrated in the tasks of image annotation and image retrieval.For the problem of deficient training classes,many researches adopts zero-shot learning methods based on embedding space,whose basic ideas are project-ing image visual features and class semantic features into a common embedding space,and recognizing test samples of new classes via the nearest neighbour rule.Domain shift and hubness are the inherent defects of this kind of method.To handle these problems,majority of existing studies introduce L1 norm or L2 norm when learning projection functions.However,the sparse estimation of L1 norm may cause underfitting of training data,while L2 norm may introduce bias in the embedding space.To this end,this dissertation proposes a novel hybrid regularization integrating the advantages of elastic net and linear discriminant analysis,which can mantain inter-class consistency and discrimination in the embedding space,thus mitigating the problems of domain shift and hubness simultaneously.An efficient synchronous optimization strategy is designed to obtain the optimal projecting functions.The proposed method has been evalu-ated through multiple benchmark image datasets,demonstrating its superiority over single regularization and many previous methods.For the problem of noisy sample labels,the current loss functions of deep neural networks(DNN)are hardly robust to noise,or are tolerant to some spe-cific noise type.Therefore,this dissertation first improves the traditional cross-entropy loss and proposes a novel rectified cross-entropy loss,which guides DNN to pay more attention to clean samples via weighting coefficients.The robustness to label noise of this loss function is proved mathematically in this dissertation.On this basis,inspired by the idea of curriculum learning and the mechanism of multi-view learning,the cross-training framework for DNN is proposed,where two synergic DNN models provides feedback information for each other about the probability of a sample containing noise,and simultane-ously update network parameters themselves.Under this framework,an online data filtration mechanism is further designed,making the framework simulta-neously optimize DNN models and filter out noisy samples.Sufficient exper-iments have been conducted on various benchmark image datasets,validating that the proposed method is robust to different noise types and noise scales.