| In recent years,with the rapid development of deep learning,the object detection algorithm based on deep learning has also been vigorously promoted,and many excellent object detection frameworks have emerged.However,most of the existing object detection algorithms rely on large-scale annotation datasets to ensure the accuracy of detection.In many practical scenes,the acquisition of a large number of annotation data is often very difficult and many manpower and material resources are consumed,which greatly limits the practical application and promotion of object detection methods.To solve this problem,more and more researchers begin to combine the few shot learning method with the object detection algorithm,so as to build a fewshot object detection technology for object detection in the case of few samples.Fewshot object detection aims to classify and locate the objects in the image through the training of a small amount of labeled data,which makes up for the shortcomings of the current object detection algorithm.It is of great research value.By analyzing some existing few-shot object detection algorithms,aiming at some problems and in order to achieve better detection results,the new few-shot object detection methods were proposed in this paper.On the one hand,based on making full use of the existing prior information,a fewshot object detection method based on Multi-Attention Mechanism and Cosine Softmax Loss was proposed.This method was improved based on the classical object detection framework Faster RCNN.Firstly,the image features were sent to CBAM(Convolution Block Attention Module)to generate attention maps,and the attention maps were multiplied with the image features for feature optimization,and then the optimized features were sent to the Attention-RPN(Region Proposal Network)combined with the idea of attention mechanism,so as to build a CBAM-Attention-RPN based on MultiAttention Mechanism,which could effectively filter out background boxes and candidate boxes of irrelevant categories.Then,the global-local relationship detector was used to further associate the support image features with the query image features by using the global relationship and local relationship,so as to obtain the candidate boxes which were more related to the object category.Finally,the classifier based on Cosine Softmax Loss was used as the classification branch of detection,which was helpful to reduce the intra-class variance and improve the detection accuracy of novel classes.For this method,relevant experiments were designed on MS COCO dataset for effectiveness verification.The experimental results show that the AP50 of this proposed method reaches 21.9,which has a certain improvement in accuracy compared with some existing few-shot object detection algorithms.On the other hand,based on metric learning,by improving the loss function,a fewshot object detection method based on Negative Margin Loss was proposed.In the field of metric learning,the conventional setting of margin to zero or positive value is conducive to increase the discrimination of the learned training features,while the negative margin loss is on the contrary.Setting the margin value to negative value reduces the distinguishability of the training category,but effectively avoids the wrong mapping of the features of the same category to multiple peaks or clusters,which improves the detection performance of few-shot object detection for novel classes.Firstly,the model with good generalization performance was trained by using largescale dataset and the object detection framework based on Negative Margin Loss.Then,the model was finetuned through a small number of samples of object categories with labels.Finally,the finetuned model was used to detect new samples of object categories.This proposed method could detect the target by learning a small number of samples.For this method,using MS COCO,PASCAL VOC and other datasets,relevant experiments are designed to verify its detection effect,and compared with some existing few-shot object detection algorithms,which have achieved better results.Few-shot object detection algorithm can reduce the demand for data in the actual scene,reduce the workload of data annotation,greatly save human,material and financial resources,and promote the landing and promotion of object detection algorithm in practical application,which has important research value and practical significance. |
