Research Based On Few-shot Object Detection Methods

Traditional object models usually require large-scale labeled samples for pre-training,which is difficult to solve related detection tasks under the condition of insufficient sample size.Therefore,in order to solve the accuracy loss caused by a small number of samples,this paper proposes a new few-shot object model.Depends on large-scale labeled data.The main work is as follows:First,a new detection model Swin T-FSOD(Swin Transformer based Few-Shot Object Detection)is proposed based on the idea of Swin Transformer.The model designs a feature learner based on the Swin Transformer and feature pyramid to extract multi-scale metafeatures containing global contextual information from the query set to detect new categories of objects.Based on the channel attention mechanism,a weight adjustment module is designed to convert the support set samples into weight coefficients with class attributes,so as to adjust the distribution of meta-features and extract the class-attribute features of the region of interest.The prediction module directly uses the detection head of the YOLOv5 framework to make predictions.Secondly,based on the Swin T-FSOD model,the SA-FSOD(Semantic Alignment based Few-Shot Object Detection)model is proposed to further improve the detection accuracy of small samples.The model proposes a new feature learner and a new weight adjustment module.In order to ensure the accuracy of feature extraction performed by the feature learner as the backbone network,a simple semantic alignment method is used to align highlevel semantic information with low-level semantic information,so that the feature extraction network no longer depends on the semantics of specific categories;use path aggregation The network performs bidirectional multi-scale fusion to solve the problem of lack of shallow location information.In the weight adjustment module,a method of adaptively adjusting the feature weight coefficient is used to enhance the interest category features in each scale meta-feature respectively.And use a task-general meta-learning initialization weight training method to further improve the generalization ability of the model.Finally,based on the above proposed few-shot object detection model,experiments are carried out on multiple data sets,and compared with the traditional few-shot detection model on various indicators.