| With the rapid growth of data volumes and computing resources,deep learning models have achieved great success in image recognition task.The success of deep learning models relies on the assumption that images are sampled independently from an identical probability distribution.However,in unconstrained scenarios,diverse images are divided into different groups with different distributions,or even different domains with distribution shift,which makes the performance of recognition system unsatisfactory.On the one hand,the distribution difference between training(source)and testing(target)data usually results in a dramatic performance drop when the model trained on training data is directly applied to recognize testing data.On the other hand,the distribution discrepancy between the different groups of training data would lead to the skewness of training process,and thus causes discriminatory decisions.Consequently,there is an increased need to guarantee the generalization and fairness for recognition systems.When recognizing the images with different distributions,mitigating the domain shift and balancing the performances between different groups are the key problems for improving the generalization and fairness of the recognition systems.Therefore,this thesis engages on proposing novel cross-domain image transformation,unsupervised domain adaptation and debiasing algorithms to endow recognition systems the capability of adaptive understanding,and applies these algorithms on some tasks,i.e.,oracle character recognition,crossdomain object classification,cross-race face recognition and fairness-aware face recognition.The main contributions can be summarized as follows.(1)To address the under-adaptation problem caused by the texture difference,a structure-texture separation network(STSN)is proposed to recognize oracle characters.STSN performs alignment and transformation based on the disentangled features.Feature disentanglement can mitigate the negative influence caused by texture,and ensure the transformed images real and diverse.Transforming images from source domain to target domain can reduce the texture difference leading to the improved target performance.To facilitate the research towards oracle character recognition,a large scale oracle character dataset called Oracle-241 is constructed.Experimental results show that STSN successfully transforms oracle data from handprinted domain to scanned domain and outperforms the traditional domain adaptation methods by 15%w.r.t target accuracy.(2)To address the negative-transfer problem caused by falsely-labeled samples,a new domain adaptation method,i.e.,Cycle Label-Consistent Networks(CLCN),is proposed to perform close-set object classification.CLCN transports label information from source domain to target domain and back to source domain,and utilizes the consistency of labels to align distribution across domains.Extensive experiments prove that it can alleviate the negative influence of falsely-labeled samples and improve the target performance.(3)To accurately cluster the unseen classes,a new domain adaptation method,i.e.,Adversarial Information Network(AIN),is proposed to perform open-set face recognition.AIN utilizes Graph Convolutional Network(GCN)to perform clustering and generate more reliable pseudo-labels.To take full advantage of these unclustered samples and mitigate the intra-domain gap,a new adversarial information loss is proposed.It iteratively modulates the positions of prototypes and pulls the features towards prototypes through the adversarial learning of mutual information,such that the separation of intra-class samples can be avoided and the discriminative features can be obtained in target domain.To facilitate the research towards face recognition across races,a large scale dataset called Transferface is presented.Compared with the baseline approaches,AIN can improve the accuracy of non-Caucasians by about 6%.(4)To address the racial bias in face recognition,a new RFW database is constructed.Experiments on RFW prove that both commercial APIs and stateof-the-art algorithms indeed suffer from racial bias,and validate that racial bias comes on both data and algorithm aspects.Furthermore,two race-aware training datasets,called Globalface and Balancedface,are provided to remove bias in training data.Finally,to mitigate the algorithmic bias,a novel Reinforcement Learning based Race Balance Network(RL-RBN)and a new Meta Balanced Network(MBN)are proposed.Based on unbiased validation set,these two methods control the learning process of network by utilizing reinforcement learning and meta learning to adjust hyperparameters,so that the feature scatter of different demographic groups can be similar leading to fair performance.Specifically,RL-RBN employs deep Q-learning to find optimal margins for non-Caucasians in which the adjustment of margins is taken as action and the balance of feature scatter of different races is taken as reward.MBN designs a new meta skewness loss and uses gradient based optimization to conduct a continuous search,which enables more suitable margin parameters.Compared with the baseline methods,RL-RBN(MBN)can reduce the standard deviation of the accuracies of different races from 1.11 to 0.73(0.58). |
