Research On Few-Shot Object Detection Technology

Object detection technology is one of the most popular technologies in the field of computer vision,both in academic research and in real-world applications.At present,most of the mainstream target detection methods based on deep learning require large-scale labeled data to realize the training process of the detection model,but in some special scenarios,such as rare animals,military security and rare diseases,it is necessary to obtain a large amount of sample data.It is difficult to label and complete the training task of the target detection model.Therefore,target detection technology based on Few-shot scenarios emerges as the times require.Based on the three directions of sample data enhancement,image feature extraction,and network model adjustment,this paper puts forward the theory and idea of target detection for Few-shot.The main work of this research is as follows:(1)A Data Augmentation method based on BIG-GAN and Mosaic is proposed.By combining Mosaic data augmentation with BIG-GAN,Few-shot data augmentation is realized through two-stage data expansion.Starting from the data source of Few-shot learning,the detection accuracy of the detection model is improved.(2)Build a Few-shot feature extraction network based on autoencoder.On the basis of VGG16 network,channel attention mechanism is added,and automatic encoder is equipped with algorithm network to realize the improvement of sample feature learning and extraction ability in the case of Few-shot.Moreover,an evaluation method based on the calculation of multiple similarity between sample images and reconstructed images is designed to quantify the feature extraction ability index.(3)A target detection algorithm based on an improved residual network is designed.By deepening the width of the residual structure on the feature extraction network,modifying the regularization,adding channels and spatial attention,and constructing a feature pool through an auto-encoder,using features Matching realizes the improvement of attention to the weighting mechanism of feature regions.In this way,the attention mechanism in the Few-shot scenario is improved,and the weight adjustment of the key area of the feature map is unreasonable,and the problem of insufficient extraction of target features and location information is eliminated,and the target detection accuracy in the Few-shot scenario can be improved from the direction of model adjustment.promote.The experimental results show that,compared with common deep learning detection algorithms,the proposed algorithm has a good effect on Few-shot data sets.The average accuracy of Few-shot data enhanced target detection is improved by about 6% on open data sets,and the image restoration similarity is improved by about 4%.The improved residual network improves the average accuracy of VOC2007 Dataset and RSOD-Dataset by 5.2%and 8% respectively.