Research On Few-Shot Learning For Image Recognition

Artificial intelligence is an important driving force for the new round of technological revolution and industrial transformation,and deep learning dominates the artificial intelligence field.The success of deep learning technology in the field of image recognition is particularly remarkable,and great achievements have been made in various application fields.However,the existing deep image recognition technologies rely on a large amount of labeled images,yet directly training the deep model with few samples is prone to overfitting.Nevertheless,acquiring sufficient labeled images requires a huge annotation cost,and for some tasks such as medical image recognition,it is impossible to collect abundant training images.As a result,few-shot learning for image recognition has become an important research problem.Although it takes a huge cost to collect adequate labeled images,it is much easy to obtain unlabeled images.Meanwhile,it is hard to directly obtain labeled data for the target task,yet a mass of labeled data may have been accrued under similar tasks.Therefore,researchers propose to explore unstinted unlabeled data or abundant of existing labeled data to assist the target task for few-shot image recognition.And we also focuse on few-shot image recognition,and according to the distribution about auxiliary data and target data,and whether the auxiliary data has annotations,this dissertation divide few-shot image recognition into three research problems:supervised few-shot image recognition,semi-supervised few-shot image recognition and unsupervised few-shot image recognition.The main contributions of this dissertation are:(1)For supervised few-shot image recognition,an algorithm by aggregating neighborhood information to enhance target feature is proposed.To aggregate the neighborhood features into the feature learning of target samples,a memory module is introduced to support the efficient acquisition of neighborhood information of target samples in the training.After obtaining the neighborhood information,in order to aggregate the neighborhood information to help learn better target sample features,the neighborhood samples are organized as a tree graph(called neighborhood tree),and the feature information of each node in the tree is iteratively aggregated into its parent node in a bottom-up way,and then the information of the whole neighborhood tree is propagated to the target feature.As the general aggregation strategy cannot determine which neighbours are more important,a supervised attention-based aggregation weight learning strategy is further proposed.Experimental results show that the proposed method can obtain more discriminative image features,and thus can obtain superior few-shot image recognition performance.For semi-supervised few-shot image recognition,a density-aware graph-based frame-work is proposed.Density information is explicitly explored to improve the feature learning and label propagation.For feature learning,a density-aware neighborhood aggregation which aggregating the neighborhood features is proposed to enhance the target features and learning the aggregation weight by considering affinity and density simultaneously.For label propagation,a density-ascending path based label propagation algorithm is introduced.As high density labeled samples can propagate more accurate labels for unlabeled samples,this dissertation further proposes to construct a density-ascending path and propagate labels from labeled samples to unlabeled samples along this path.Finally,the density-aware feature aggregation and label propagation algorithms are integrated into a unified framework.The experimental results show that the framework can not only improve the feature learning,but also efficiently propagate labels for unlabeled samples,thus obtain better semi-supervised classification perfor-mance.For unsupervised few-shot image recognition,this dissertation proposes target-aware unsupervised pretraining and eigen-finetuning.Firstly,we observe that the clustering performance of target samples in the feature space is crucial for few-shot learning of target task.Moreover,we observes that the contrastive loss based unsupervised learning share similar spirit with clustering which both push similar samples to be more closer and conversely,dissimilar images to have a far distance.Based on above observations,a target-aware unsupervised pretraining strategy is proposed,in which the unlabeled target samples together with the unlabeled source samples are involved in unsupervised pretraining.Further,the improved clustering is of great value for identifying the most representative samples("eigen-samples")for users to label,and in return,continued finetuning with the labeled eigen-samples further improves the clustering.Thus,this dissertation proposes eigen-finetuning by leveraging the co-evolution of clustering and eigen-samples in the finetuning.Experimental results show that the proposed approach can effectively boost the few-shot transfer ability of unsupervised pretraining,and can achieve comparable performance with methods based on supervised pretraining.