Research On Small Sample Target Detection Technology Based On Deep Learning

Small sample target detection technology has significant application value in the military and security fields,it can effectively reduce the training time,and can save a lot of marked data,so as to improve the accuracy and efficiency of detection.Therefore,this paper effectively improves the detection accuracy in small sample scenarios through three aspects: data enhancement,meta-learning model improvement and transfer learning algorithm strategy improvement.The main contents of the research are as follows:(1)A new data enhancement method is proposed,which can enhance small sample data effectively.Aircraft fake samples generated in generative adversarial network are combined with pre-selected image background to improve the accuracy of detection model.This method can not only guarantee the quality of samples,but also effectively expand the data set of small samples.Through the verification experiment on the NWPU VHR-10 dataset,the data enhancement method proposed in this paper can effectively improve the accuracy of the data,with an average improvement rate of up to 6%.(2)A multi-scale feature fusion technique and channel reweighting method based on meta-learning were constructed,and a spatial-channel attention mechanism was introduced to enhance the saliency of the target in the image,so as to effectively extract the early features of the detection network.After the feature pyramid network(FPN)of Faster R-CNN detection framework,pixel aggregation network(PAN)is added.Through the fusion of transposition convolution and other technologies with low-level location features and high-level semantic features,better information fusion is achieved,which improves the identification accuracy of small targets,and provides better detection performance in dense target scenarios.The method of Euclidean distance calculation by prototype network is used to solve the problem of candidate region classification.Compared with the traditional model,the proposed improved model improves the detection accuracy by 1.91%(3)A small sample target detection algorithm based on transfer learning is proposed to balance the number of regional suggestion boxes of base class and new class by improving the pre-training stage and fine-tuning stage in the process of transfer learning.The refined method PRA is proposed.The method proposed in this paper can effectively improve the detection accuracy of small samples based on the two-stage target detection model.Compared with the mainstream algorithm,PRA algorithm improves the detection accuracy by 2～6%.It is found that compared with traditional detection algorithms,the data enhancement method proposed in this paper can improve the average accuracy of different categories without significantly reducing the detection speed.The multi-scale feature fusion and channel-space reweighting mechanism based on meta-learning proposed in this paper can effectively improve the target detection accuracy of the Faster-CNN model.The PRA algorithm proposed in this paper can improve the detection accuracy of new classes by adjusting the number distribution of regional suggestion boxes.