Research On Few Shot Object Detection Method Based On Semi-supervised Learning

Object detection is one of the core tasks in the field of computer vision,and has extremely important application and research significance.However,the current object detection methods rely on large-scale labeled data training to achieve satisfactory detection accuracy.The production of large-scale labeled data is very labor-intensive and material resources.In some scenarios,sufficient labeled data cannot even be collected.Therefore,a small amount of labeled data is used to guide and Semisupervised learning methods trained with massive unlabeled data have attracted much attention,among which pseudo-label semi-supervised learning methods are widely used.At present,there are some difficulties in the few shot object detection method based on pseudo-label semi-supervised learning: how to select the correct labeled data for model training;how to update model to generate higher-quality pseudo-label data.Focusing on the above difficulties,the main research work of the thesis is as follows:(1)Aiming at the problem that selecting correctly labeled pseudo-label data in the pseudo-label semi-supervised learning method,a pseudo-label object detection method based on classification uncertainty(Soft Teacher-CUC-EMAD)is proposed.The Soft Teacher-CUC-EMAD method uses the classification uncertainty method on the basis of the Soft Teacher method to calculate the uncertainty of the pseudo-label classification results to screen high-quality pseudo-label data.The lower the uncertainty,the more reliable the classification results.Uncertainty is added to the classification loss function of pseudo-label data as a weight to reduce the negative impact of highuncertainty pseudo-labels on the model;finally,the EMAD method is used to update the Teacher model,reducing the similarity between the parameters of the Student model and Teacher model,making the consistency regularization method to work better.The verification experiments were carried out when the labeled data accounted for 1%,5%,and 10% of the training set,respectively.Compared with the Soft Teacher method,the detection accuracy of the Soft Teacher-CUC-EMAD method increased by 1.4%,1.2%,and 1.7%,respectively.It is verified that the Soft Teacher-CUC-EMAD method improves the detection accuracy of pseudo-label target detection.(2)Aiming at the problem that the widely used EMA model update method in the pseudo-label semi-supervised learning method causes the similarity of the Teacher model and the Student model to cause insufficient model learning,the MPLTU model update method is proposed,and a combined MPLTU pseudo-label object detection model(UTML)is constructed.The UTML model uses the MPLTU method to update the Teacher model on the basis of the Unbiased Teacher model,which improves the independence of the Teacher model and alleviates the similarity problem of the dual models;a consistency regularization loss function is added to enhance the model’s antinoise ability;finally,the training strategy is adjusted and modified to improve the model detection accuracy;at the same time,the UTML model is combined with the classification uncertainty method proposed in(1)to obtain the UTMC model(UTML with CUCU),which is used to verify the positive effect of the classification uncertainty method..Experiments were carried out under the labeled data of 1%,5% and 10% of train2017.Compared with the Unbiased Teacher model,the detection accuracy of the UTML model was improved by 0.7%,0.5% and 0.5%,respectively,which verified that the UTML method improved the detection accuracy of pseudo-labeled objects detection accuracy: Compared with the UTML model,the detection accuracy of the UTMC model is increased by 0.2%,0.1% and 0.1% respectively,which verifies that the classification uncertainty has a certain improvement in the detection performance of the UTML model.(3)Studying object detection in remote sensing images under pseudo-label semisupervised learning method,a pseudo-label object detection method based on YOLOS(UTMC-YOLOS)is proposed.The UTMC-YOLOS method is based on the UTMC method proposed in(2),which combines the classification uncertainty and the MPLTU model update method,and selects YOLOS,which is more suitable for remote sensing object detection,as the Teacher model and Student model.The uncertainty method is modified.Comparing experiments on the training set composed of the labeled data set NWPU VHR-10 and the unlabeled data set HRRSD,the detection accuracy of the UTMC method reaches 63.1%,compared with the YOLOS model improved by 0.9%,which verifies that the UTMC method has a good detection effect in the field of remote sensing images.