Supernova Object Detection Method Based On Faster R-CNN

Supernova search is of great significance in exploring the history of the expansion of the universe and studying cosmology and astronomy.The traditional method of supernova search is mainly achieved through the image processing for early detection,and the advanced users checks the detection results later,but this method has the problems of complex algorithm design,large environmental impact,and high labor costs.Deep learning has gained widespread attention in recent years,by optimizing and training the network,deep learning can automatically learn the intrinsic characteristics of data.For specific scenarios,high accuracy model can be obtained without learning professional knowledge.In this paper,an algorithm for supernova object detection based on the Faster R-CNN is proposed because of the complex image background,small object,imbalance between positive and negative samples,which leads to the poor image contrast,weak feature expression,and poor detection performance.The main contents are:(1)Aiming at the problem that the image contrast is poor,this paper firstly uses the methods of image flipping and rotating to increase the data set,and then synthesizes the new image,the old image and the difference image to improve the image contrast,reduce the difficulty of feature learning.Experimental results show that,after image synthesis,the average value of customized evaluation index Score is increased by 30.06%,F1 also improved significantly.(2)In order to improve the capability of feature expression,this paper firstly uses the deep residual network to extract the features of the composite image,and the high-level features were successively fused with the low-level features to build a regional proposal network pyramid,so that each layer of the network has strong semantic information,and optimizes the design of anchors and the feature extractor for region of interest.Using the feature extracted in this way to detect the object significantly improves the accuracy.(3)In order to solve the problem of the imbalance between positive and negative samples,this paper use the online hard example mining method to select the hard samples in the model training,and continuously added the hard samples into the training set,which improved the judgment ability of the hard samples,solved the problem of imbalance between positive and negative samples.(4)In order to improve the detection accuracy,this paper uses multi-scale test method in image detection to improve the perception ability of image scale.The improved algorithm was compared with the original Faster R-CNN,SSD,RFCN algorithm,and the customized evaluation index Score and F1 were higher.It was verified on the PASCAL VOC 2012 data set,and the experiment showed that the m AP improved by 9.32% compared with the original Faster R-CNN algorithm,and the generalization ability was better.