Research On Small Sample Target Detection Algorithm Based On Deep Learning

At present,the object detection algorithms and technologies have been initially applied in intelligent robot vision and obstacle detection,intelligent license plate detection,intelligent face detection,and object detection in network image monitoring,and have become an important and popular research topic in the field of computer vision.In recent years,with the continuous research of deep learning algorithms,the object detection algorithm based on deep learning has also made great progress,but it needs to use a large amount of data with label information for training.In reality,it is difficult to obtain large-scale data with labeled information.So how to use data with only a little label information to learn and train a model with a certain generalization ability is a difficult problem.we makes following contributions.1.To solve the problem of few-shot target detection,we combined faster-RCNN with the transfer learning in few-shot learning to construct a few-shot object detection network structure.Among them,the first stage was to use a large number of base class samples to train the network.The second stage used the parameters obtained in the first stage to initialize the network and used base data and novel data with a small number of samples to fine-tune the model.2.To solve the problem of insufficient sample information in few-shot object detection and poor detection results for small targets,we used the feature pyramid structure and proposes a few-shot target detection algorithm based on feature fusion.By analyzing the characteristics of each layer of the faster-RCNN and the pyramid structure,we makeuped the feature layer fusion structure.The algorithm combined the deep and shallow features of the sample image to make full use of the strong semantics of the deep features and the high resolution features of the shallow features to improve the characterization ability of the samples.Experimental results proved that this method could effectively improve the accuracy of small target detection in few-shot object detection.3.Aiming at the problem of imbalance in the number of difficult and easy samples,we improved the classification loss function in the few-shot object detection algorithm by using focal loss.Keep the ratio of difficult and easy samples unchanged,the percentage of difficult and easy samples’ loss to the classification loss is adjusted by hyperparameter,to enlarge the useful information’ percentage,and to reduce the useless information of percentage to the classification loss.Experimental results proved that this method effectively improved the accuracy of few-shot object detection.