Key Technologies Of Image Recognition Under Limited Supervision

In recent years,with the rapid growth of computing resources and training data,ma-chine learning and deep learning models have achieved promising performance in various applications,such as computer vision,natural language processing,and biological informa-tion computing.Especially in the field of image recognition and medical image analysis,deep learning models under fully supervision have achieved similar or even better recog-nition accuracy compared with human experts.However,the impressive achievements of existing machine learning and deep learning models are highly dependent on large-scale and fine-grained data annotations.Collecting and labeling a large amount of training data for each new task is expensive or even unavailable in many real application scenarios.There-fore,how to obtain satisfactory model performance with limited supervision is of great importance in machine learning and deep learning community.This article mainly focuses on image recognition in limited supervision scenarios,which aims to answers the follow-ing questions:(1)How to use a large amount of source domain data and a small amount of target domain data to improve the model performance on the target task.(2)How to train a high performance model on the target task with a large number of annotated source data and unannotated target data.(3)How to train a fine-grained image recognition model with only weakly supervised(image-level)annotations.Main contributions of this article are summarized as follows:Parameter Transfer Extreme Learning Machine Based on Projective Model.As a widely used classification model,extreme learning machine(ELM)can not fit well when the given task has limited training samples.In order to reduce the dependence of ELM model on the target data,a parameter transfer extreme learning machine based on the pro-jection model is proposed,which is termed as PTELM.The proposed PTELM uses a linear transformation matrix to bridge the source model parameters and the target model parame-ters.By jointly learning the linear transformation matrix and the classification hyperplane parameters,the PTELM is able to effectively transfer the information of the source data to the target model,which prevents the target model from over-fitting.Joint Domain Matching and Classification for Cross-domain Adaptation via ELM.The PTELM model only adapts the classification hyperplane,which can not handle the sit-uation where the source and target domain have significant difference in feature space.Therefore,we further proposes a cross-domain ELM by joint domain alignment and classi-fication.In this work,by aligning the marginal distribution and conditional distribution of the source and target data simultaneously,the distribution discrepancy between the source and target domain can be reduced more effectively.In addition,we introduce a l21 regu-larization to the model parameters,which encourages the model to select more informa-tive features for knowledge transfer.Finally,the proposed JDMC model integrates joint distribution alignment,feature selection and output space adaptation into a unified ELM framework.Joint Domain Matching and Discriminative Feature Learning for Unsupervised Domain Adaptation.Unsupervised domain adaptation aims to use a large amount of source domain data to label unlabeled target domain data.Existing methods only focus on learning domain-invariant features by minimizing the discrepancy between the source and target domain data.Due to lack of supervision in the target domain,the leaned deep features in the target domain is non-discriminative,which leads to poor generalization performance for the target domain data.In this work,a novel self-similarity consistency metric is intro-duced to minimize the domain distribution discrepancy.In addition,we propose to learn more discriminative deep features by improving the intra-class compactness and inter-class separability of deep features,which significantly improves the generalization performance of the model on the target data.Higher-order Moment Matching for Unsupervised Domain Adaptation.For un-supervised domain adaptation problem,most existing discrepancy-based methods are de-signed to match the second-order or lower moments,which however,have limited expres-sion of statistical characteristic for non-Gaussian distributions.In this work,we propose a Higher-order Moment Matching(HoMM)method,and further extend the HoMM into reproducing kernel Hilbert spaces(RKHS).In particular,the proposed HoMM can per-form arbitrary-order moment matching,we show that the first-order HoMM is equivalent to Maximum Mean Discrepancy and the second-order HoMM is equivalent to Correlation Alignment.Moreover,HoMM(order≥ 3)is expected to perform fine-grained domain alignment as higher-order statistics can approximate more complex,non-Gaussian distri-butions.Extensive experiments are conducted,showing that our proposed HoMM consis-tently outperforms the existing domain discrepancy-based methods by a large margin.Bone Age Assessment with Weakly Supervised Learning.Bone age estimation(BAA)is a fine-grained image recognition problem,which can be used to diagnose en-docrine and metabolic disorders during child development.In order to make use of local fine-grained information to obtain better BAA performance,existing methods rely on pro-viding additional local annotations.In this work,we propose a weakly supervised BAA method,which is able to obtain local information with only image-level annotation.Our model can find the hand region,the most discriminative region(the carpal bones),and the next most discriminative region(the metacarpal bones)according to the learned attention maps.Besides,instead of taking BAA as a general regression task,which is suboptimal due to the label ambiguity problem in the age label space,we propose using joint age dis-tribution learning and expectation regression,which makes use of the ordinal relationship among hand images with different individual ages and leads to more robust age estimation.Extensive experiments show that our method achieves competitive results compared with existing state-of-the-art methods that require manual annotation.In summary,this article proposes several learning models based on limited supervi-sion,aiming to solve the problem of limited training data or limited data annotation in areas image recognition and computer-aided diagnosis.The proposed model has achieved con-siderable performance on multiple tasks such as cross-domain digital recognition,cross-domain natural image recognition,cross-domain CovID-19 diagnosis,and fine-grained bone age estimation.