Research On Few-shot Object Detection Based On Deep Learning

In recent years,by introducing deep learning techniques and convolutional neural networks,researchers in the field of computer vision have achieved very good results on general object detection.However,as a data-dependent technique,the effectiveness of deep learning is closely related to the size of the data involved in training.When the size of the data available for training is small,the training samples often fail to represent the distribution of targets in realistic environments and do not fully reflect the differences in shape,pose and texture that exist between similar targets.As a result,existing models are severely over-fitted and lack basic generalisation capabilities.In real-life situations,however,insufficient samples are common.In areas such as security,healthcare and content censorship,the cost of labeling samples for specific targets is high,and some samples are naturally insufficient,causing considerable problems for the practical application of AI in such scenarios.To address these issues,this paper first explores a suitable implementation of a few-shot classification network.It is then combined with a two-stage object detection framework to propose a new distance metric learning based few-shot object detection network,which mainly includes the following research points.(1)A distance metric learning-based method for few-sample classification is investigated.The method achieves classification by mapping images into the classification space and then following the nearest neighbor idea.Firstly,a convolutional neural network is used to extract and map the features of the image while modelling all the categories under the classification space,and then the distance between the image and all the category representations under the space is calculated and mapped to 0?1,representing the probability that the image belongs to each category.The experimental results show that the classification method based on distance metric learning in this paper can achieve the classification of few-shot targets more accurately,with good accuracy and recall.(2)A distance metric learning-based target detection method for few-sample targets is proposed.The method introduces the previously studied few-shot classification network,replaces the second stage of the classification network with a few-shot classification module based on the Faster RCNN,and constructs a multi-branch design to sample and extract features from both the support set images and the training set images in training,and then models the background targets separately.Again,a category representation for each category is constructed in the classification space and the ROI regions are classified accordingly.The method will also use a two-stage training strategy to train the RPN module of the network and the target regression network in the second stage to improve the network's ability to localise the few-shot targets and ultimately achieve few-shot object detection.Compared with existing methods,this method is able to construct multiple modalities for the few-shot category,reflecting as much as possible its different morphological features under the same category,while the detection speed is not affected.The experimental results show that the proposed distance metric-based object detection network has relatively good performance.