Research On Few Shot Object Detection And Recognition

Object detection and recognition is a basic task and a hot research issue in the field of computer vision.At present,most of the research on this direction is based on a large number of labeled data and high performance computing power.Then,when the training data is scarce,the training accuracy of the model will decline sharply,because of the over-fitting,which will lead to poor and generalization ability.In order to solve these problems,this paper studies the problem from two perspectives of improving network representation ability and loss function.The specific content is as follows:Existing few-shot object detection methods mainly design new training strategies for existing object detection networks,but ignore the importance of network representation ability.Therefore,starting from improving network structure,this paper proposes a new model named HOSENET.In this model,higher-order semantic enhancement modules(such as second-order pooling module)are introduced into the forward process of the network.Firstly,the network is trained with the data of the base class,then the parameters of the network are fine-tuned with the novel class data.Finally,the objects of the new class are detected with the trained model.Relevant experiments have been conducted on both Pascal VOC and MS COCO datasets,and the experiments show that,without affecting the detection accuracy of the base class,the detection accuracy of our method in the new class is significantly better than that of other methods.Then,aiming at the practical application of Mei school gesture recognition in Peking Opera performing art,this paper proposes a novel method of Mei school gesture recognition.This method is based on few shot learning,using the training framework of meta-learning,and using the loss function based on vector space distance and cosine distance.Firstly,53 gesture pictures of Mei school and gesture pictures with numbers 0?10 were collected to form a dataset,and the dataset was preprocessed and divided.Then,a network was constructed to extract the features of digital gesture images,and a loss function was designed to optimize the model parameters.Finally,the network is fine-tuned with the new class(Mei school gesture)data,and the fine-tuned network model is used for Mei school gesture recognition.The experimental results verify the effectiveness of the method,which provides a strong technical support for the inheritance and development of Mei school gesture.