Research On The Building And Learning Problems Of Label Distribution In The Label Ambiguity Representation Of Images

In machine vision,resembling semantic labels and image content variation result in the label ambiguity of images,which can be represented by the label distribution that assigns a sample image multiple labels with different intensities.On the basis of the single truth label of training samples,the process of building label distributions to capture label ambiguity can be interpreted as label smoothing reinforcement,which can boost the feature representation and classification performance for convolutional neural networks.The ideal label ambiguity representation can be given by human annotation,and learning the map from sample features to the given label distributions is called label distribution learning,which can be used to infer the label ambiguity about new samples.It is theoretically and practically significant to represent label ambiguity for promoting network performance conditioned on minor samples,enhancing network adaption against complicated scenes,and explaining sample polysemia.Focusing on the label ambiguity representation of images,this thesis proposes three methods to build label distributions and a label distribution model.The thesis mainly research the following contents:(1)Aiming at the problem that the difference of label ambiguity degrees among samples are underestimated,aleatoric uncertainty is exploited to measure label ambiguity degree,an aleatoric uncertainty-based adaptive label distribution method is proposed.Firstly,The Bayesian neural network is constructed to learn the aleatoric uncertainty of samples.Then,the label ambiguity degrees of samples is measured according to the uncertainty,and a large smoothing variance is used to build label distributions if the uncertainty is high.The proposed method is applied to the task of pose estimation of Sichuan pepper in the pepper picking robot,and the task of age estimation task on a public face dataset.The experiments show that,the proposed method sufficiently distinguishes the difference of label ambiguity degrees among samples,reduces the confusion among resembling labels,and is universal in promoting the classification performance for continuous labels.(2)Aiming at the problem that the relationship among discrete labels is unknown and thus it impossible to build label distributions,label relationship is measured through probabilistic topic models,a label topic embedding-based label distribution method is proposed.Firstly,the topic model is exploited to compute the topic distributions of samples as well as the topic embeddings of labels.The topics reflect latent label attributions,and those labels having close embedding values on the same topic are treated as neighboring labels.Then,label intensities are smoothed to neighboring labels,and the label smoothing results on different topics are integrated by using the sample topic distributions as weights,which leads to the label distributions that jointly represent label relationship and sample semantics.The experiments on public remote sensing images and professional component images show that,the proposed method stably captures the label ambiguity of samples,promotes network feature and classification performance,and the promotion is especially obvious conditioned on minor samples.(3)Aiming at the problem that current subspace learning methods emphasize discrimination ability,and thus it is suboptimal to use the current methods to represent label ambiguity,a new subspace learning model is defined,a sample affinity graph-based label distribution method is proposed.Firstly,a subspace learning model is defined to find sample neighboring relations that can reflect label ambiguity,which consists of a sparse regularizing item to capture local similarity and a global regularizing item to handle label relationship.These two regularizing items enforce the solved neighboring samples share similar appearance and correlated labels.After that,label distributions are obtained through the maximum a posterior estimation on the class frequencies of the neighboring samples.The experiments on remote sensing images and component images shows that,the label distributions can adapt to image content variation and includes little label noise,the label ambiguity of samples is effectively captured.The proposed method significantly promotes the feature representation and classification performance of networks.(4)Aiming at the problem that the technologies of clustering-based grouping and low rank constraint adopted in current label distribution learning methods can hardly exploit local label relationship,nonnegative matrix factorization is exploited to capture local label relationship,a label distribution learning model based on nonnegative dictionary is proposed.Firstly,graph-regularized nonnegative matrix factorization is employed to extract nonnegative dictionary from the label distribution matrix of training set.The dictionary is used to represent local label relationship,which prevents clustering-based grouping from destructing sample data structure.Then,the model forces the outputted label distributions can be reconstructed by the dictionary,and thus the model outputs are guaranteed to satisfy local label relationship,but not constrained by low rank constraints.The experiment on human expression and natural scenes show that,the nonnegative dictionary stably captures local label relationship,and the prediction ability for label distributions achieved by the proposed model often surpasses that of current methods.