Research On Image Classification And Object Detection Based On Few-Shot Learning

Deep learning has attracted a lot of attention recently.People once again anticipate deep learning as a result of the tremendous advancements in its research and applications.Deep learning technology has taken over as the primary approach to problem-solving in the field of computer vision research.Deep learning,however,needs a lot of labeled data to train on.It is very expensive to obtain and label a large amount of data.It is impossible to acquire a lot of data in some unique situations.Deep learning must therefore be tailored to the investigation of few-shot data size.At present,the study of few-shot learning is still in its infancy,and there is still a big gap between the performance of few-shot learning and that of large sample learning.Therefore,this paper studies the image classification problem and object detection problem under the condition of few-shot to improve their performance.First,from the perspective of data augment to alleviate the problem of few samples,three data augment methods based on simple style transfer,SSMixup,SSCutMix,and SSCutout,are proposed.Introducing style transfer into data augment can increase the diversity of augmented data,and thus alleviate the problem of small samples to a certain extent.At the same time,these three methods use the style transfer method based on transformation,which basically does not introduce additional computation during style transfer,thus ensuring the training efficiency of the model.Experiments show that these three methods can achieve high performance under the conditions of large samples and small samples.Secondly,a few-shot image classification method based on self-supervised task pretraining is proposed,which uses self-supervised task to improve the generalization performance of the model in the pretraining stage.The model pretraining is divided into three stages.In the first stage,contrastive learning tasks are used for training to promote the model to learn invariant knowledge.In the second stage,rotational prediction tasks and classification tasks are used to promote the model to learn equivariant knowledge and discriminated knowledge.In the third stage,self distillation methods are used to improve the performance of the model.Experiments show that this method can achieve very advanced performance.Finally,aiming at the problem of poor generalization of few-shot object detection task model,a few-shot object detection method based on Cascade RCNN is proposed.Improve the performance of few-shot object detection from model pretraining,model architecture,data augment,model inferring and other aspects.This method greatly improves the generalization performance of few-shot object detection model.