| Image cognitive model based on the deep neural network has been widely used in various felds and achieved great success with its powerful feature extraction ability and generalization ability.The core problem of image recognition,image segmentation,and many other tasks related to image cognition is how to extract discriminative representation accurately and effciently.The representation learning method extracts discriminative representation by using the rich information contained within data and has achieved remarkable results in the felds of unsupervised,semi-supervised,and fully supervised learning.However,the increasingly wide range of tasks and more practical applications have presented many serious challenges to representation learning.Among them,how to further relax the applicable conditions of representation learning,reduce the cost of learning and improve the effciency of learning are urgent problems to be solved.Especially in the application of medical images,the particularity of medical images and the complexity of clinical situations pose unique challenges to representation learning.Aiming at the key problems in the mechanism,algorithm,and application of representation learning,this dissertation explores the learning mechanism of discriminative representation based on traditional machine learning theory.The algorithm for extracting discriminative representation is studied based on two key points(expression space optimization and feature decorrelation).Taking typical challenges in medical image application as an example,a discriminative representation extraction model for practical application is designed.Firstly,at the mechanism level,this dissertation starts from the traditional theoretical methods of machine learning to study the mechanism of learning discriminative representation.Based on this,it discusses the key points for learning discriminative representations.Secondly,at the algorithm level,this dissertation proposes a probability distribution-based discriminative representation learning method to solve the problem that learning discriminative representation depends on prior knowledge of data.To solve the problems of low learning effciency and requiring large computing resources,a discriminative representation learning method based on indirect feature decorrelation is proposed.Finally,based on the application of medical images,this dissertation proposes a segmentation method based on locally discriminative representation to solve the problem that it is diffcult to learn discriminative representation from samples with unclear semantic features.A domain adaptation method based on feature channels is proposed to solve the problem that model transferring is diffcult based on a small amount of data.The innovations and major contributions of this dissertation include the following aspects:1.An explanation of the learning mechanism of discriminative representation is proposed.Taking classical contrastive learning as an example,this dissertation studies the mechanism of learning discriminative representation by exploring the correlation between contrast learning and spectral clustering.It explores the key points of learning discriminative representation.The representation learning method based on contrastive thought has achieved good results in many tasks and verifed the efectiveness of contrastive learning.The essence of the success of such methods can be seen as achieving the extraction of discriminative representations.However,few studies have explained why the contrastive-based method can extract discriminative representations.This dissertation studies the mechanism of extracting discriminative representations by taking contrastive learning as an example.Based on the observation that both contrast learning and spectral clustering are optimized based on a similar matrix,this dissertation analyzes the learning mechanism of discriminative representation by discussing the correlation between contrast learning and spectral clustering.In order to prove the intrinsic relationship between contrastive learning and spectral clustering,it frst proves that the optimization of contrastive learning’s loss function can indirectly optimize the objective function of spectral clustering.At the same time,it shows that contrast learning can naturally promote feature de-correlation.Therefore,it points out that,under certain conditions,contrastive learning is essentially similar to spectral clustering to learn discriminative representation with small intra-class and large intra-class distances.In addition,the infuence of the two key points of optimal contrastive learning loss function and feature de-correlation on discriminative representation learning is discussed.Finally,the internal relationship between contrast learning and spectral clustering and the infuence of the two key points on discriminative representation learning is verifed.2.An optimization method of representation space based on probability distribution is proposed to realize the extraction of highly discriminative representation that can be directly used in clustering without any prior knowledge.By removing the dependence on prior knowledge in the learning process,the conditions of discriminative representation learning are relaxed,and the applicability of discriminative representation in multi-grained environments is improved.One of the core goals of unsupervised representation learning is to learn discriminative representation without human supervision.In general,discriminative representation with small intra-class and large inter-class distances can be efectively clustered by basic clustering methods.However,most of the existing methods need to use data prior knowledge to learn highly discriminative representations that can be used for clustering.However,such prior knowledge may not be readily available in real-world scenarios.In addition,when the prior knowledge changes,the representation extraction network needs to be retrained before it can be used in new scenarios.This dissertation proposes a contrastive learning method based on probability distribution,which can learn a highly discriminative representation for clustering without using prior knowledge of data sets.This dissertation encodes the input data into a multivariate Gaussian distribution.Samples are drawn from the distribution,therefore realizing the contrastive learning from two perspectives:probability distribution and random sample.Compared with coding samples into representation vectors,coding samples into probabilistic distribution enables each sample to cover more representation space,thus forcing the network to optimize the arrangement of representation space.The method proposed in this dissertation is based on probability distribution,which strengthens the constraint on representation space so that the network can learn discriminative representation without prior knowledge.The representation learned by this method can obtain a good unsupervised clustering result through simple K-means clustering and surpasses the previous SOTA method by 9% on the classifcation accuracy of CIFAR-10.Finally,the efectiveness of the proposed method is verifed by experimental results.At the same time,the multi-grained clustering is taken as an example to demonstrate the stability of the method when the prior knowledge of data changes.3.A discriminative representation learning method based on indirect feature decorrelation simplifed the optimization objective of the loss function and improved the effciency of learning discriminative representation.And the bi-vector-based contrastive learning method is proposed,which realizes the small batch-based contrastive learning and reduces the demand for computing resources.In order to obtain highly discriminative representations,integrating feature decorrelation techniques into the learning framework is an empirically efective idea.However,these methods usually de-correlate features in an additional feature space,which makes the optimization process need to be carried out simultaneously in the representation and feature spaces,reducing learning effciency.In addition,the representation learning method based on contrastive thought usually requires a large number of optimization,which makes the model’s training need a lot of computational resources.Therefore,improving the learning effciency of discriminative representations and reducing the demand for computing resources is an urgent problem to be solved.To improve the learning effciency,this dissertation proposes a feature de-correlation strategy based on small-batch contrastive learning,which uses the correlation between the dimension of feature space and batch size to indirectly de-correlation features.A bi-vector-based contrastive learning method is proposed to realize contrastive learning based on small batch sizes and reduce the demand for computing resources.Finally,the efectiveness of the proposed method is verifed by experiments and surpasses the previous SOTA method by 7% on the classifcation accuracy of CIFAR-10.The proposed method can reduce the demand for resources in contrastive learning and improve the effciency of the expression of learning separability.4.A locally discriminative representation learning method is proposed to realize the extraction of discriminative representation of samples with vague semantic features.A caries segmentation model without additional mask training was designed to improve the accuracy of caries screening and reduce the workload of labeling.In many medical image application scenarios,it is diffcult to distinguish the semantic features of the target and some background structures in the image,so it is diffcult to extract the discriminative representation.As a typical disease that is diffcult to distinguish the lesion and background,caries dramatically afects the general population’s health.Caries segmentation based on the panoramic image is an efective method for automatic caries screening.In this dissertation,the caries segmentation task on panoramic images is taken as an example to study how to extract the discriminative representation from the data with vague semantic features.Caries segmentation based on panoramic images is challenging.The more efective multi-phase approach relies on additional annotations,while the simple end-to-end approach is less efective.In this dissertation,a new end-to-end network is proposed to improve the accuracy of caries screening based on panoramic images by exploring how to extract locally discriminative representations.To learn the local separable representation,this dissertation conducts fne-grained clustering for the representation.It proposes a probabilistic inference module so that the coarse-grained annotation can provide supervision signals for fne-grained clustering.An adaptive multi-head mapping module is proposed to map locally discriminative representations to coarse-grained annotations,mapping local representations with diferent criteria.The efectiveness of the proposed method was verifed by experiments based on the caries segmentation data set of panoramic images and surpassed the previous SOTA method by 1% on Dice.The proposed method is as efective as the multistage method requiring additional masks and has a more favorable performance for caries screening.5.a feature-based domain adaptation method is proposed to achieve common optimization based on representation space and feature space,improve the effciency of domain adaptive in sample utilization,and reduce the demand for data volume.Taking the children’s caries examination as an example,the segmentation model transferring based on a small amount of data was realized.The acquisition cost of medical data is very high.The effcient reuse of existing data and network models to reduce the requirement of new task data is signifcant in medical image applications.Domain adaptation can reduce the need for data by transferring existing models.Screening of children’s panoramic dental images is a small data scene.In the cases of children’s caries,severe caries in young children,and multiple caries,panoramic dental imaging is an efective way to detect caries.The segmentation technology for children’s caries can help doctors improve the effciency of screening caries.However,the number of children’s dental implants is relatively small compared with adult dental implants.Therefore,this dissertation takes children’s caries segmentation task as an example to study how to use small data for domain adaptation.Based on the fact that there are few panoramic dental images in children,this dissertation studies how to transfer the caries segmentation model based on permanent adult teeth to the caries segmentation task of children’s deciduous teeth with a small amount of data and proposes an adaptive domain method based on feature space.This method uses domain adversarial adaptive learning in both representation and feature spaces to learn discriminative representation with domain invariance.A feature domain adaptive method based on similarity screening and a normalized strategy based on feature modulus screening is proposed to achieve a stable and effcient feature domain adaptive method.A contrastive learning strategy based on ordering augmentation is proposed to constrain the consistency of feature vectors to eliminate the infuence of random ordering of samples in a batch.This strategy can carry out natural feature decorrelation,strengthen the constraint on discriminative quality of representation,and improve the transfer effciency of feature space.Based on the segmented caries data set of children’s panoramic dental images,this dissertation verifes the validity of the proposed method,which can use a small amount of data to transfer the model and reduce the dependence on labeling and data.This dissertation proposes a feature space domain adaptation method based on similarity selection to achieve a stable and effcient adaptive feature space domain.To improve the infuence of samples on the adaptive feature channels,a method based on feature orthogonalization is proposed to encourage the limited target domain samples to spread to all feature channels.The efectiveness of the proposed method is experimentally verifed on a dataset of caries segmentation based on children’s panoramic dental images and surpassed the previous SOTA method by 10% on Dice.The proposed method can transfer models based on a small amount of data and reduce the dependence on doctors’ annotations. |
