Object Detection By Autonomous Learning Based On Few-shot Metric Learning

Object detection is widely used in the fields of automatic driving,security video surveillance and so on.It is an important research direction and hotspot of computer vision.In recent years,with the continuous development of deep learning,the object detection method based on Convolutional Neural Network(CNN)has gradually replaced the traditional object detection method and become the mainstream of the current research in this field.However,this object detection method usually relies on a large number of labeled training samples.In practical application,it is not easy to obtain a large number of labeled samples,which often requires long-term accumulation or high cost,and limits the application of such detection methods to a large extent.Some researchers have proposed a object detection method based on few-shot learning,trying to train a detection model with good generalization performance through a small number of labeled samples.However,due to the lack of sample number and poor sample diversity within the class,it is a difficult problem to obtain a detection model that meets the requirements from limited data.As a distance learning method,deep metric learning can obtain the distribution relationship of different samples in characteristic space,and then measure the similarity between each sample.In order to adapt to the problem that diversity of the same class of data is poor,deep metric learning in limited samples to extract the characteristics of a large number of distribution relations,after the effective learning of these distribution relationships,it not only can accurately distinguish the differences between different samples,but also can extract transferable in-class deformation from similar samples.This paper proposes a object detection method by autonomous learning based on fewshot metric learning,and then measures the similarity between the object region and the instance by constructing a deep metric network that can extract multi-scale features;On this basis,the model refinement strategy was proposed,and the training samples were further expanded through the automatic mining of approximate instances,and then the metric model was refined iteratively for several times to continuously accumulate and discover new knowledge,so as to improve its generalization ability.The innovation points of this paper can be briefly summarized in the following four aspects:(1)In order to enrich the feature representation,a deep metric network based on SPP feature pyramid pooling is proposed.The network can extract visual features of samples at different scales,and fix the output of the pooling layer.On this basis,the output is combined with the global features extracted from the "support" network and the "query" network to further improve the richness and robustness of feature expression.(2)In order to synchronize sample expansion and model refinement training,a model autonomous learning strategy is proposed.According to the similarity between the object region and the instance sample,the strategy independently mines the neighboring region of the central instance in the image,so that the model can gradually discover the implicit visual information of the object in the process of refining,so as to continuously reduce the gap between the few-shot instance and the central instance.(3)In order to weaken the influence of extreme instances and make the selected central sample more representative,a instance query method based on regional confidence expression is proposed.In the process of classification,the method measures the average similarity between the image region and each instance sample,and obtains the category confidence representation of the region.The classification problem is transformed into the problem of comparing the distance between the image region and each instance center.(4)On the basis of the comparative loss,the loss function is optimized,besides introducing the maximum interval control,regularization penalty term is also introduced on the basis of the loss function to form the structural risk function,so as to control the influence of the inconsistency of different feature transformation.The proposed method was validated on PASCAL VOC2007 and PASCAL VOC2012 datasets.The algorithm performs well in object detection by weakening the restriction of label information,which proves the effectiveness of the self-learning method based on fewshot metric in object detection.It can get rid of the limitation of sample size to a large extent and has certain application value.